searchengines-web/MeiliSearch
searchengines-web/MeiliSearch
1
0
Fork 0
mirror of https://github.com/meilisearch/MeiliSearch synced 2024-12-26 14:40:05 +01:00
Code Issues Releases Wiki Activity

Merge branch 'main' into release-v1.7.5-tmp

Browse Source
This commit is contained in:
Tamo 2024-04-09 14:27:48 +02:00
commit c26d356a35
93 changed files with 6574 additions and 2093 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
.github/ISSUE_TEMPLATE/sprint_issue.md vendored
View File

@ -2,14 +2,13 @@
name: New sprint issue
about: ⚠️ Should only be used by the engine team ⚠️
title: ''
labels: ''
labels: 'missing usage in PRD, impacts docs'
assignees: ''
---
Related product team resources: [PRD]() (_internal only_)
Related product discussion:
Related spec: WIP
## Motivation
@ -21,11 +20,7 @@ Related spec: WIP
## TODO
<!---Feel free to adapt this list with more technical/product steps-->
- [ ] Release a prototype
- [ ] If prototype validated, merge changes into `main`
- [ ] Update the spec
<!---If necessary, create a list with technical/product steps-->
### Reminders when modifying the Setting API

2
.github/workflows/bench-pr.yml vendored
View File

@ -43,4 +43,4 @@ jobs:
- name: Run benchmarks on PR ${{ github.event.issue.id }}
run: |
cargo xtask bench --api-key "${{ secrets.BENCHMARK_API_KEY }}" --dashboard-url "${{ vars.BENCHMARK_DASHBOARD_URL }}" --reason "[Comment](${{ github.event.comment.url }}) on [#${{github.event.issue.id}}](${{ github.event.issue.url }})" -- ${{ steps.command.outputs.command-arguments }}
cargo xtask bench --api-key "${{ secrets.BENCHMARK_API_KEY }}" --dashboard-url "${{ vars.BENCHMARK_DASHBOARD_URL }}" --reason "[Comment](${{ github.event.comment.html_url }}) on [#${{ github.event.issue.number }}](${{ github.event.issue.html_url }})" -- ${{ steps.command.outputs.command-arguments }}

38
.github/workflows/milestone-workflow.yml vendored
View File

@ -110,6 +110,44 @@ jobs:
--milestone $MILESTONE_VERSION \
--assignee curquiza
create-update-version-issue:
needs: get-release-version
# Create the update-version issue even if the release is a patch release
if: github.event.action == 'created'
runs-on: ubuntu-latest
env:
ISSUE_TEMPLATE: issue-template.md
steps:
- uses: actions/checkout@v3
- name: Download the issue template
run: curl -s https://raw.githubusercontent.com/meilisearch/engine-team/main/issue-templates/update-version-issue.md > $ISSUE_TEMPLATE
- name: Create the issue
run: |
gh issue create \
--title "Update version in Cargo.toml for $MILESTONE_VERSION" \
--label 'maintenance' \
--body-file $ISSUE_TEMPLATE \
--milestone $MILESTONE_VERSION
create-update-openapi-issue:
needs: get-release-version
# Create the openAPI issue if the release is not only a patch release
if: github.event.action == 'created' && needs.get-release-version.outputs.is-patch == 'false'
runs-on: ubuntu-latest
env:
ISSUE_TEMPLATE: issue-template.md
steps:
- uses: actions/checkout@v3
- name: Download the issue template
run: curl -s https://raw.githubusercontent.com/meilisearch/engine-team/main/issue-templates/update-openapi-issue.md > $ISSUE_TEMPLATE
- name: Create the issue
run: |
gh issue create \
--title "Update Open API file for $MILESTONE_VERSION" \
--label 'maintenance' \
--body-file $ISSUE_TEMPLATE \
--milestone $MILESTONE_VERSION
# ----------------
# MILESTONE CLOSED
# ----------------

362
BENCHMARKS.md Normal file
View File

@ -0,0 +1,362 @@
# Benchmarks
Currently this repository hosts two kinds of benchmarks:
1. The older "milli benchmarks", that use [criterion](https://github.com/bheisler/criterion.rs) and live in the "benchmarks" directory.
2. The newer "bench" that are workload-based and so split between the [`workloads`](./workloads/) directory and the [`xtask::bench`](./xtask/src/bench/) module.
This document describes the newer "bench" benchmarks. For more details on the "milli benchmarks", see [benchmarks/README.md](./benchmarks/README.md).
## Design philosophy for the benchmarks
The newer "bench" benchmarks are **integration** benchmarks, in the sense that they spawn an actual Meilisearch server and measure its performance end-to-end, including HTTP request overhead.
Since this is prone to fluctuating, the benchmarks regain a bit of precision by measuring the runtime of the individual spans using the [logging machinery](./CONTRIBUTING.md#logging) of Meilisearch.
A span roughly translates to a function call. The benchmark runner collects all the spans by name using the [logs route](https://github.com/orgs/meilisearch/discussions/721) and sums their runtime. The processed results are then sent to the [benchmark dashboard](https://bench.meilisearch.dev), which is in charge of storing and presenting the data.
## Running the benchmarks
Benchmarks can run locally or in CI.
### Locally
#### With a local benchmark dashboard
The benchmarks dashboard lives in its [own repository](https://github.com/meilisearch/benchboard). We provide binaries for Ubuntu/Debian, but you can build from source for other platforms (MacOS should work as it was developed under that platform).
Run the `benchboard` binary to create a fresh database of results. By default it will serve the results and the API to gather results on `http://localhost:9001`.
From the Meilisearch repository, you can then run benchmarks with:
```sh
cargo xtask bench -- workloads/my_workload_1.json ..
```
This command will build and run Meilisearch locally on port 7700, so make sure that this port is available.
To run benchmarks on a different commit, just use the usual git command to get back to the desired commit.
#### Without a local benchmark dashboard
To work with the raw results, you can also skip using a local benchmark dashboard.
Run:
```sh
cargo xtask bench --no-dashboard -- workloads/my_workload_1.json workloads/my_workload_2.json ..
```
For processing the results, look at [Looking at benchmark results/Without dashboard](#without-dashboard).
### In CI
We have dedicated runners to run workloads on CI. Currently, there are three ways of running the CI:
1. Automatically, on every push to `main`.
2. Manually, by clicking the [`Run workflow`](https://github.com/meilisearch/meilisearch/actions/workflows/bench-manual.yml) button and specifying the target reference (tag, commit or branch) as well as one or multiple workloads to run. The workloads must exist in the Meilisearch repository (conventionally, in the [`workloads`](./workloads/) directory) on the target reference. Globbing (e.g., `workloads/*.json`) works.
3. Manually on a PR, by posting a comment containing a `/bench` command, followed by one or multiple workloads to run. Globbing works. The workloads must exist in the Meilisearch repository in the branch of the PR.
```
/bench workloads/movies*.json /hackernews_1M.json
```
## Looking at benchmark results
### On the dashboard
Results are available on the global dashboard used by CI at <https://bench.meilisearch.dev> or on your [local dashboard](#with-a-local-benchmark-dashboard).
The dashboard homepage presents three sections:
1. The latest invocations (a call to `cargo xtask bench`, either local or by CI) with their reason (generally set to some helpful link in CI) and their status.
2. The latest workloads ran on `main`.
3. The latest workloads ran on other references.
By default, the workload shows the total runtime delta with the latest applicable commit on `main`. The latest applicable commit is the latest commit for workload invocations that do not originate on `main`, and the latest previous commit for workload invocations that originate on `main`.
You can explicitly request a detailed comparison by span with the `main` branch, the branch or origin, or any previous commit, by clicking the links at the bottom of the workload invocation.
In the detailed comparison view, the spans are sorted by improvements, regressions, stable (no statistically significant change) and unstable (the span runtime is comparable to its standard deviation).
You can click on the name of any span to get a box plot comparing the target commit with multiple commits of the selected branch.
### Without dashboard
After the workloads are done running, the reports will live in the Meilisearch repository, in the `bench/reports` directory (by default).
You can then convert these reports into other formats.
- To [Firefox profiler](https://profiler.firefox.com) format. Run:
```sh
cd bench/reports
cargo run --release --bin trace-to-firefox -- my_workload_1-0-trace.json
```
You can then upload the resulting `firefox-my_workload_1-0-trace.json` file to the online profiler.
## Designing benchmark workloads
Benchmark workloads conventionally live in the `workloads` directory of the Meilisearch repository.
They are JSON files with the following structure (comments are not actually supported, to make your own, remove them or copy some existing workload file):
```jsonc
{
// Name of the workload. Must be unique to the workload, as it will be used to group results on the dashboard.
"name": "hackernews.ndjson_1M,no-threads",
// Number of consecutive runs of the commands that should be performed.
// Each run uses a fresh instance of Meilisearch and a fresh database.
// Each run produces its own report file.
"run_count": 3,
// List of arguments to add to the Meilisearch command line.
"extra_cli_args": ["--max-indexing-threads=1"],
// List of named assets that can be used in the commands.
"assets": {
// name of the asset.
// Must be unique at the workload level.
// For better results, the same asset (same sha256) should have the same name accross workloads.
// Having multiple assets with the same name and distinct hashes is supported accross workloads,
// but will lead to superfluous downloads.
//
// Assets are stored in the `bench/assets/` directory by default.
"hackernews-100_000.ndjson": {
// If the assets exists in the local filesystem (Meilisearch repository or for your local workloads)
// Its file path can be specified here.
// `null` if the asset should be downloaded from a remote location.
"local_location": null,
// URL of the remote location where the asset can be downloaded.
// Use the `--assets-key` of the runner to pass an API key in the `Authorization: Bearer` header of the download requests.
// `null` if the asset should be imported from a local location.
// if both local and remote locations are specified, then the local one is tried first, then the remote one
// if the file is locally missing or its hash differs.
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-100_000.ndjson",
// SHA256 of the asset.
// Optional, the `sha256` of the asset will be displayed during a run of the workload if it is missing.
// If present, the hash of the asset in the `bench/assets/` directory will be compared against this hash before
// running the workload. If the hashes differ, the asset will be downloaded anew.
"sha256": "60ecd23485d560edbd90d9ca31f0e6dba1455422f2a44e402600fbb5f7f1b213",
// Optional, one of "Auto", "Json", "NdJson" or "Raw".
// If missing, assumed to be "Auto".
// If "Auto", the format will be determined from the extension in the asset name.
"format": "NdJson"
},
"hackernews-200_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-200_000.ndjson",
"sha256": "785b0271fdb47cba574fab617d5d332276b835c05dd86e4a95251cf7892a1685"
},
"hackernews-300_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-300_000.ndjson",
"sha256": "de73c7154652eddfaf69cdc3b2f824d5c452f095f40a20a1c97bb1b5c4d80ab2"
},
"hackernews-400_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-400_000.ndjson",
"sha256": "c1b00a24689110f366447e434c201c086d6f456d54ed1c4995894102794d8fe7"
},
"hackernews-500_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-500_000.ndjson",
"sha256": "ae98f9dbef8193d750e3e2dbb6a91648941a1edca5f6e82c143e7996f4840083"
},
"hackernews-600_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-600_000.ndjson",
"sha256": "b495fdc72c4a944801f786400f22076ab99186bee9699f67cbab2f21f5b74dbe"
},
"hackernews-700_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-700_000.ndjson",
"sha256": "4b2c63974f3dabaa4954e3d4598b48324d03c522321ac05b0d583f36cb78a28b"
},
"hackernews-800_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-800_000.ndjson",
"sha256": "cb7b6afe0e6caa1be111be256821bc63b0771b2a0e1fad95af7aaeeffd7ba546"
},
"hackernews-900_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-900_000.ndjson",
"sha256": "e1154ddcd398f1c867758a93db5bcb21a07b9e55530c188a2917fdef332d3ba9"
},
"hackernews-1_000_000.ndjson": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/hackernews/hackernews-1_000_000.ndjson",
"sha256": "27e25efd0b68b159b8b21350d9af76938710cb29ce0393fa71b41c4f3c630ffe"
}
},
// Core of the workload.
// A list of commands to run sequentially.
// A command is a request to the Meilisearch instance that is executed while the profiling runs.
"commands": [
{
// Meilisearch route to call. `http://localhost:7700/` will be prepended.
"route": "indexes/movies/settings",
// HTTP method to call.
"method": "PATCH",
// If applicable, body of the request.
// Optional, if missing, the body will be empty.
"body": {
// One of "empty", "inline" or "asset".
// If using "empty", you can skip the entire "body" key.
"inline": {
// when "inline" is used, the body is the JSON object that is the value of the `"inline"` key.
"displayedAttributes": [
"title",
"by",
"score",
"time"
],
"searchableAttributes": [
"title"
],
"filterableAttributes": [
"by"
],
"sortableAttributes": [
"score",
"time"
]
}
},
// Whether to wait before running the next request.
// One of:
// - DontWait: run the next command without waiting the response to this one.
// - WaitForResponse: run the next command as soon as the response from the server is received.
// - WaitForTask: run the next command once **all** the Meilisearch tasks created up to now have finished processing.
"synchronous": "DontWait"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
// When using "asset", use the name of an asset as value to use the content of that asset as body.
// the content type is derived of the format of the asset:
// "NdJson" => "application/x-ndjson"
// "Json" => "application/json"
// "Raw" => "application/octet-stream"
// See [AssetFormat::to_content_type](https://github.com/meilisearch/meilisearch/blob/7b670a4afadb132ac4a01b6403108700501a391d/xtask/src/bench/assets.rs#L30)
// for details and up-to-date list.
"asset": "hackernews-100_000.ndjson"
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-200_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-300_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-400_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-500_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-600_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-700_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-800_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-900_000.ndjson"
},
"synchronous": "WaitForResponse"
},
{
"route": "indexes/movies/documents",
"method": "POST",
"body": {
"asset": "hackernews-1_000_000.ndjson"
},
"synchronous": "WaitForTask"
}
]
}
```
### Adding new assets
Assets reside in our DigitalOcean S3 space. Assuming you have team access to the DigitalOcean S3 space:
1. go to <https://cloud.digitalocean.com/spaces/milli-benchmarks?i=d1c552&path=bench%2Fdatasets%2F>
2. upload your dataset:
1. if your dataset is a single file, upload that single file using the "upload" button,
2. otherwise, create a folder using the "create folder" button, then inside that folder upload your individual files.
## Upgrading `https://bench.meilisearch.dev`
The URL of the server is in our password manager (look for "benchboard").
1. Make the needed modifications on the [benchboard repository](https://github.com/meilisearch/benchboard) and merge them to main.
2. Publish a new release to produce the Ubuntu/Debian binary.
3. Download the binary locally, send it to the server:
```
scp -6 ~/Downloads/benchboard root@\[<ipv6-address>\]:/bench/new-benchboard
```
Note that the ipv6 must be between escaped square brackets for SCP.
4. SSH to the server:
```
ssh root@<ipv6-address>
```
Note the ipv6 must **NOT** be between escaped square brackets for SSH 🥲
5. On the server, set the correct permissions for the new binary:
```
chown bench:bench /bench/new-benchboard
chmod 700 /bench/new-benchboard
```
6. On the server, move the new binary to the location of the running binary (if unsure, start by making a backup of the running binary):
```
mv /bench/{new-,}benchboard
```
7. Restart the benchboard service.
```
systemctl restart benchboard
```
8. Check that the service runs correctly.
```
systemctl status benchboard
```
9. Check the availability of the service by going to <https://bench.meilisearch.dev> on your browser.

26
CONTRIBUTING.md
View File

@ -4,7 +4,7 @@ First, thank you for contributing to Meilisearch! The goal of this document is t
Remember that there are many ways to contribute other than writing code: writing [tutorials or blog posts](https://github.com/meilisearch/awesome-meilisearch), improving [the documentation](https://github.com/meilisearch/documentation), submitting [bug reports](https://github.com/meilisearch/meilisearch/issues/new?assignees=&labels=&template=bug_report.md&title=) and [feature requests](https://github.com/meilisearch/product/discussions/categories/feedback-feature-proposal)...
The code in this repository is only concerned with managing multiple indexes, handling the update store, and exposing an HTTP API. Search and indexation are the domain of our core engine, [`milli`](https://github.com/meilisearch/milli), while tokenization is handled by [our `charabia` library](https://github.com/meilisearch/charabia/).
Meilisearch can manage multiple indexes, handle the update store, and expose an HTTP API. Search and indexation are the domain of our core engine, [`milli`](https://github.com/meilisearch/meilisearch/tree/main/milli), while tokenization is handled by [our `charabia` library](https://github.com/meilisearch/charabia/).
If Meilisearch does not offer optimized support for your language, please consider contributing to `charabia` by following the [CONTRIBUTING.md file](https://github.com/meilisearch/charabia/blob/main/CONTRIBUTING.md) and integrating your intended normalizer/segmenter.
@ -81,6 +81,30 @@ Meilisearch follows the [cargo xtask](https://github.com/matklad/cargo-xtask) wo
Run `cargo xtask --help` from the root of the repository to find out what is available.
### Logging
Meilisearch uses [`tracing`](https://lib.rs/crates/tracing) for logging purposes. Tracing logs are structured and can be displayed as JSON to the end user, so prefer passing arguments as fields rather than interpolating them in the message.
Refer to the [documentation](https://docs.rs/tracing/0.1.40/tracing/index.html#using-the-macros) for the syntax of the spans and events.
Logging spans are used for 3 distinct purposes:
1. Regular logging
2. Profiling
3. Benchmarking
As a result, the spans should follow some rules:
- They should not be put on functions that are called too often. That is because opening and closing a span causes some overhead. For regular logging, avoid putting spans on functions that are taking less than a few hundred nanoseconds. For profiling or benchmarking, avoid putting spans on functions that are taking less than a few microseconds.
- For profiling and benchmarking, use the `TRACE` level.
- For profiling and benchmarking, use the following `target` prefixes:
- `indexing::` for spans meant when profiling the indexing operations.
- `search::` for spans meant when profiling the search operations.
### Benchmarking
See [BENCHMARKS.md](./BENCHMARKS.md)
## Git Guidelines
### Git Branches

705
Cargo.lock generated
View File

File diff suppressed because it is too large Load Diff

2
Cargo.toml
View File

@ -21,7 +21,7 @@ members = [
]
[workspace.package]
version = "1.7.5"
version = "1.8.0"
authors = [
"Quentin de Quelen <quentin@dequelen.me>",
"Clément Renault <clement@meilisearch.com>",

910
assets/grafana-dashboard.json
View File

File diff suppressed because it is too large Load Diff

1
dump/src/lib.rs
View File

@ -277,6 +277,7 @@ pub(crate) mod test {
}),
pagination: Setting::NotSet,
embedders: Setting::NotSet,
search_cutoff_ms: Setting::NotSet,
_kind: std::marker::PhantomData,
};
settings.check()

1
dump/src/reader/compat/v5_to_v6.rs
View File

@ -379,6 +379,7 @@ impl<T> From<v5::Settings<T>> for v6::Settings<v6::Unchecked> {
v5::Setting::NotSet => v6::Setting::NotSet,
},
embedders: v6::Setting::NotSet,
search_cutoff_ms: v6::Setting::NotSet,
_kind: std::marker::PhantomData,
}
}

2
dump/src/reader/v2/updates.rs
View File

@ -61,7 +61,7 @@ pub enum IndexDocumentsMethod {
#[cfg_attr(test, derive(serde::Serialize))]
#[non_exhaustive]
pub enum UpdateFormat {
/// The given update is a real **comma seperated** CSV with headers on the first line.
/// The given update is a real **comma separated** CSV with headers on the first line.
Csv,
/// The given update is a JSON array with documents inside.
Json,

2
dump/src/writer.rs
View File

@ -219,7 +219,7 @@ pub(crate) mod test {
fn _create_directory_hierarchy(dir: &Path, depth: usize) -> String {
let mut ret = String::new();
// the entries are not guarenteed to be returned in the same order thus we need to sort them.
// the entries are not guaranteed to be returned in the same order thus we need to sort them.
let mut entries =
fs::read_dir(dir).unwrap().collect::<std::result::Result<Vec<_>, _>>().unwrap();

2
filter-parser/src/value.rs
View File

@ -42,7 +42,7 @@ fn quoted_by(quote: char, input: Span) -> IResult<Token> {
)));
}
}
// if it was preceeded by a `\` or if it was anything else we can continue to advance
// if it was preceded by a `\` or if it was anything else we can continue to advance
}
Ok((

2
index-scheduler/src/autobatcher.rs
View File

@ -870,7 +870,7 @@ mod tests {
debug_snapshot!(autobatch_from(false,None, [doc_imp(UpdateDocuments, false, None), settings(false), idx_del()]), @"Some((IndexDeletion { ids: [0, 2, 1] }, false))");
debug_snapshot!(autobatch_from(false,None, [doc_imp(ReplaceDocuments,false, None), settings(false), doc_clr(), idx_del()]), @"Some((IndexDeletion { ids: [1, 3, 0, 2] }, false))");
debug_snapshot!(autobatch_from(false,None, [doc_imp(UpdateDocuments, false, None), settings(false), doc_clr(), idx_del()]), @"Some((IndexDeletion { ids: [1, 3, 0, 2] }, false))");
// The third and final case is when the first task doesn't create an index but is directly followed by a task creating an index. In this case we can't batch whith what
// The third and final case is when the first task doesn't create an index but is directly followed by a task creating an index. In this case we can't batch whit what
// follows because we first need to process the erronous batch.
debug_snapshot!(autobatch_from(false,None, [doc_imp(ReplaceDocuments,false, None), settings(true), idx_del()]), @"Some((DocumentOperation { method: ReplaceDocuments, allow_index_creation: false, primary_key: None, operation_ids: [0] }, false))");
debug_snapshot!(autobatch_from(false,None, [doc_imp(UpdateDocuments, false, None), settings(true), idx_del()]), @"Some((DocumentOperation { method: UpdateDocuments, allow_index_creation: false, primary_key: None, operation_ids: [0] }, false))");

6
index-scheduler/src/batch.rs
View File

@ -920,7 +920,11 @@ impl IndexScheduler {
}
// 3.2. Dump the settings
let settings = meilisearch_types::settings::settings(index, &rtxn)?;
let settings = meilisearch_types::settings::settings(
index,
&rtxn,
meilisearch_types::settings::SecretPolicy::RevealSecrets,
)?;
index_dumper.settings(&settings)?;
Ok(())
})?;

66
index-scheduler/src/lib.rs
View File

@ -1301,8 +1301,8 @@ impl IndexScheduler {
wtxn.commit().map_err(Error::HeedTransaction)?;
// Once the tasks are commited, we should delete all the update files associated ASAP to avoid leaking files in case of a restart
tracing::debug!("Deleting the upadate files");
// Once the tasks are committed, we should delete all the update files associated ASAP to avoid leaking files in case of a restart
tracing::debug!("Deleting the update files");
//We take one read transaction **per thread**. Then, every thread is going to pull out new IDs from the roaring bitmap with the help of an atomic shared index into the bitmap
let idx = AtomicU32::new(0);
@ -1332,7 +1332,7 @@ impl IndexScheduler {
Ok(TickOutcome::TickAgain(processed_tasks))
}
/// Once the tasks changes have been commited we must send all the tasks that were updated to our webhook if there is one.
/// Once the tasks changes have been committed we must send all the tasks that were updated to our webhook if there is one.
fn notify_webhook(&self, updated: &RoaringBitmap) -> Result<()> {
if let Some(ref url) = self.webhook_url {
struct TaskReader<'a, 'b> {
@ -3028,6 +3028,66 @@ mod tests {
snapshot!(serde_json::to_string_pretty(&documents).unwrap(), name: "documents");
}
#[test]
fn test_settings_update() {
use meilisearch_types::settings::{Settings, Unchecked};
use milli::update::Setting;
let (index_scheduler, mut handle) = IndexScheduler::test(true, vec![]);
let mut new_settings: Box<Settings<Unchecked>> = Box::default();
let mut embedders = BTreeMap::default();
let embedding_settings = milli::vector::settings::EmbeddingSettings {
source: Setting::Set(milli::vector::settings::EmbedderSource::Rest),
api_key: Setting::Set(S("My super secret")),
url: Setting::Set(S("http://localhost:7777")),
..Default::default()
};
embedders.insert(S("default"), Setting::Set(embedding_settings));
new_settings.embedders = Setting::Set(embedders);
index_scheduler
.register(
KindWithContent::SettingsUpdate {
index_uid: S("doggos"),
new_settings,
is_deletion: false,
allow_index_creation: true,
},
None,
false,
)
.unwrap();
index_scheduler.assert_internally_consistent();
snapshot!(snapshot_index_scheduler(&index_scheduler), name: "after_registering_settings_task");
{
let rtxn = index_scheduler.read_txn().unwrap();
let task = index_scheduler.get_task(&rtxn, 0).unwrap().unwrap();
let task = meilisearch_types::task_view::TaskView::from_task(&task);
insta::assert_json_snapshot!(task.details);
}
handle.advance_n_successful_batches(1);
snapshot!(snapshot_index_scheduler(&index_scheduler), name: "settings_update_processed");
{
let rtxn = index_scheduler.read_txn().unwrap();
let task = index_scheduler.get_task(&rtxn, 0).unwrap().unwrap();
let task = meilisearch_types::task_view::TaskView::from_task(&task);
insta::assert_json_snapshot!(task.details);
}
// has everything being pushed successfully in milli?
let index = index_scheduler.index("doggos").unwrap();
let rtxn = index.read_txn().unwrap();
let configs = index.embedding_configs(&rtxn).unwrap();
let (_, embedding_config) = configs.first().unwrap();
insta::assert_json_snapshot!(embedding_config.embedder_options);
}
#[test]
fn test_document_replace_without_autobatching() {
let (index_scheduler, mut handle) = IndexScheduler::test(false, vec![]);

13
index-scheduler/src/snapshots/index_scheduler__tests__settings_update-2.snap Normal file
View File

@ -0,0 +1,13 @@
---
source: index-scheduler/src/lib.rs
expression: task.details
---
{
"embedders": {
"default": {
"source": "rest",
"apiKey": "MyXXXX...",
"url": "http://localhost:7777"
}
}
}

23
index-scheduler/src/snapshots/index_scheduler__tests__settings_update-3.snap Normal file
View File

@ -0,0 +1,23 @@
---
source: index-scheduler/src/lib.rs
expression: embedding_config.embedder_options
---
{
"Rest": {
"api_key": "My super secret",
"distribution": null,
"dimensions": null,
"url": "http://localhost:7777",
"query": null,
"input_field": [
"input"
],
"path_to_embeddings": [
"data"
],
"embedding_object": [
"embedding"
],
"input_type": "text"
}
}

13
index-scheduler/src/snapshots/index_scheduler__tests__settings_update.snap Normal file
View File

@ -0,0 +1,13 @@
---
source: index-scheduler/src/lib.rs
expression: task.details
---
{
"embedders": {
"default": {
"source": "rest",
"apiKey": "MyXXXX...",
"url": "http://localhost:7777"
}
}
}

36
index-scheduler/src/snapshots/lib.rs/test_settings_update/after_registering_settings_task.snap Normal file
View File

@ -0,0 +1,36 @@
---
source: index-scheduler/src/lib.rs
---
### Autobatching Enabled = true
### Processing Tasks:
[]
----------------------------------------------------------------------
### All Tasks:
0 {uid: 0, status: enqueued, details: { settings: Settings { displayed_attributes: NotSet, searchable_attributes: NotSet, filterable_attributes: NotSet, sortable_attributes: NotSet, ranking_rules: NotSet, stop_words: NotSet, non_separator_tokens: NotSet, separator_tokens: NotSet, dictionary: NotSet, synonyms: NotSet, distinct_attribute: NotSet, proximity_precision: NotSet, typo_tolerance: NotSet, faceting: NotSet, pagination: NotSet, embedders: Set({"default": Set(EmbeddingSettings { source: Set(Rest), model: NotSet, revision: NotSet, api_key: Set("My super secret"), dimensions: NotSet, document_template: NotSet, url: Set("http://localhost:7777"), query: NotSet, input_field: NotSet, path_to_embeddings: NotSet, embedding_object: NotSet, input_type: NotSet, distribution: NotSet })}), search_cutoff_ms: NotSet, _kind: PhantomData<meilisearch_types::settings::Unchecked> } }, kind: SettingsUpdate { index_uid: "doggos", new_settings: Settings { displayed_attributes: NotSet, searchable_attributes: NotSet, filterable_attributes: NotSet, sortable_attributes: NotSet, ranking_rules: NotSet, stop_words: NotSet, non_separator_tokens: NotSet, separator_tokens: NotSet, dictionary: NotSet, synonyms: NotSet, distinct_attribute: NotSet, proximity_precision: NotSet, typo_tolerance: NotSet, faceting: NotSet, pagination: NotSet, embedders: Set({"default": Set(EmbeddingSettings { source: Set(Rest), model: NotSet, revision: NotSet, api_key: Set("My super secret"), dimensions: NotSet, document_template: NotSet, url: Set("http://localhost:7777"), query: NotSet, input_field: NotSet, path_to_embeddings: NotSet, embedding_object: NotSet, input_type: NotSet, distribution: NotSet })}), search_cutoff_ms: NotSet, _kind: PhantomData<meilisearch_types::settings::Unchecked> }, is_deletion: false, allow_index_creation: true }}
----------------------------------------------------------------------
### Status:
enqueued [0,]
----------------------------------------------------------------------
### Kind:
"settingsUpdate" [0,]
----------------------------------------------------------------------
### Index Tasks:
doggos [0,]
----------------------------------------------------------------------
### Index Mapper:
----------------------------------------------------------------------
### Canceled By:
----------------------------------------------------------------------
### Enqueued At:
[timestamp] [0,]
----------------------------------------------------------------------
### Started At:
----------------------------------------------------------------------
### Finished At:
----------------------------------------------------------------------
### File Store:
----------------------------------------------------------------------

40
index-scheduler/src/snapshots/lib.rs/test_settings_update/settings_update_processed.snap Normal file
View File

@ -0,0 +1,40 @@
---
source: index-scheduler/src/lib.rs
---
### Autobatching Enabled = true
### Processing Tasks:
[]
----------------------------------------------------------------------
### All Tasks:
0 {uid: 0, status: succeeded, details: { settings: Settings { displayed_attributes: NotSet, searchable_attributes: NotSet, filterable_attributes: NotSet, sortable_attributes: NotSet, ranking_rules: NotSet, stop_words: NotSet, non_separator_tokens: NotSet, separator_tokens: NotSet, dictionary: NotSet, synonyms: NotSet, distinct_attribute: NotSet, proximity_precision: NotSet, typo_tolerance: NotSet, faceting: NotSet, pagination: NotSet, embedders: Set({"default": Set(EmbeddingSettings { source: Set(Rest), model: NotSet, revision: NotSet, api_key: Set("My super secret"), dimensions: NotSet, document_template: NotSet, url: Set("http://localhost:7777"), query: NotSet, input_field: NotSet, path_to_embeddings: NotSet, embedding_object: NotSet, input_type: NotSet, distribution: NotSet })}), search_cutoff_ms: NotSet, _kind: PhantomData<meilisearch_types::settings::Unchecked> } }, kind: SettingsUpdate { index_uid: "doggos", new_settings: Settings { displayed_attributes: NotSet, searchable_attributes: NotSet, filterable_attributes: NotSet, sortable_attributes: NotSet, ranking_rules: NotSet, stop_words: NotSet, non_separator_tokens: NotSet, separator_tokens: NotSet, dictionary: NotSet, synonyms: NotSet, distinct_attribute: NotSet, proximity_precision: NotSet, typo_tolerance: NotSet, faceting: NotSet, pagination: NotSet, embedders: Set({"default": Set(EmbeddingSettings { source: Set(Rest), model: NotSet, revision: NotSet, api_key: Set("My super secret"), dimensions: NotSet, document_template: NotSet, url: Set("http://localhost:7777"), query: NotSet, input_field: NotSet, path_to_embeddings: NotSet, embedding_object: NotSet, input_type: NotSet, distribution: NotSet })}), search_cutoff_ms: NotSet, _kind: PhantomData<meilisearch_types::settings::Unchecked> }, is_deletion: false, allow_index_creation: true }}
----------------------------------------------------------------------
### Status:
enqueued []
succeeded [0,]
----------------------------------------------------------------------
### Kind:
"settingsUpdate" [0,]
----------------------------------------------------------------------
### Index Tasks:
doggos [0,]
----------------------------------------------------------------------
### Index Mapper:
doggos: { number_of_documents: 0, field_distribution: {} }
----------------------------------------------------------------------
### Canceled By:
----------------------------------------------------------------------
### Enqueued At:
[timestamp] [0,]
----------------------------------------------------------------------
### Started At:
[timestamp] [0,]
----------------------------------------------------------------------
### Finished At:
[timestamp] [0,]
----------------------------------------------------------------------
### File Store:
----------------------------------------------------------------------

2
meilisearch-types/Cargo.toml
View File

@ -11,7 +11,7 @@ edition.workspace = true
license.workspace = true
[dependencies]
actix-web = { version = "4.4.1", default-features = false }
actix-web = { version = "4.5.1", default-features = false }
anyhow = "1.0.79"
convert_case = "0.6.0"
csv = "1.3.0"

13
meilisearch-types/src/error.rs
View File

@ -2,6 +2,7 @@ use std::{fmt, io};
use actix_web::http::StatusCode;
use actix_web::{self as aweb, HttpResponseBuilder};
use aweb::http::header;
use aweb::rt::task::JoinError;
use convert_case::Casing;
use milli::heed::{Error as HeedError, MdbError};
@ -56,7 +57,14 @@ where
impl aweb::error::ResponseError for ResponseError {
fn error_response(&self) -> aweb::HttpResponse {
let json = serde_json::to_vec(self).unwrap();
HttpResponseBuilder::new(self.status_code()).content_type("application/json").body(json)
let mut builder = HttpResponseBuilder::new(self.status_code());
builder.content_type("application/json");
if self.code == StatusCode::SERVICE_UNAVAILABLE {
builder.insert_header((header::RETRY_AFTER, "10"));
}
builder.body(json)
}
fn status_code(&self) -> StatusCode {
@ -259,6 +267,7 @@ InvalidSettingsProximityPrecision , InvalidRequest , BAD_REQUEST ;
InvalidSettingsFaceting , InvalidRequest , BAD_REQUEST ;
InvalidSettingsFilterableAttributes , InvalidRequest , BAD_REQUEST ;
InvalidSettingsPagination , InvalidRequest , BAD_REQUEST ;
InvalidSettingsSearchCutoffMs , InvalidRequest , BAD_REQUEST ;
InvalidSettingsEmbedders , InvalidRequest , BAD_REQUEST ;
InvalidSettingsRankingRules , InvalidRequest , BAD_REQUEST ;
InvalidSettingsSearchableAttributes , InvalidRequest , BAD_REQUEST ;
@ -304,6 +313,7 @@ MissingSwapIndexes , InvalidRequest , BAD_REQUEST ;
MissingTaskFilters , InvalidRequest , BAD_REQUEST ;
NoSpaceLeftOnDevice , System , UNPROCESSABLE_ENTITY;
PayloadTooLarge , InvalidRequest , PAYLOAD_TOO_LARGE ;
TooManySearchRequests , System , SERVICE_UNAVAILABLE ;
TaskNotFound , InvalidRequest , NOT_FOUND ;
TooManyOpenFiles , System , UNPROCESSABLE_ENTITY ;
TooManyVectors , InvalidRequest , BAD_REQUEST ;
@ -352,6 +362,7 @@ impl ErrorCode for milli::Error {
| UserError::InvalidOpenAiModelDimensions { .. }
| UserError::InvalidOpenAiModelDimensionsMax { .. }
| UserError::InvalidSettingsDimensions { .. }
| UserError::InvalidUrl { .. }
| UserError::InvalidPrompt(_) => Code::InvalidSettingsEmbedders,
UserError::TooManyEmbedders(_) => Code::InvalidSettingsEmbedders,
UserError::InvalidPromptForEmbeddings(..) => Code::InvalidSettingsEmbedders,

129
meilisearch-types/src/settings.rs
View File

@ -202,12 +202,52 @@ pub struct Settings<T> {
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default, error = DeserrJsonError<InvalidSettingsEmbedders>)]
pub embedders: Setting<BTreeMap<String, Setting<milli::vector::settings::EmbeddingSettings>>>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default, error = DeserrJsonError<InvalidSettingsSearchCutoffMs>)]
pub search_cutoff_ms: Setting<u64>,
#[serde(skip)]
#[deserr(skip)]
pub _kind: PhantomData<T>,
}
impl<T> Settings<T> {
pub fn hide_secrets(&mut self) {
let Setting::Set(embedders) = &mut self.embedders else {
return;
};
for mut embedder in embedders.values_mut() {
let Setting::Set(embedder) = &mut embedder else {
continue;
};
let Setting::Set(api_key) = &mut embedder.api_key else {
continue;
};
Self::hide_secret(api_key);
}
}
fn hide_secret(secret: &mut String) {
match secret.len() {
x if x < 10 => {
secret.replace_range(.., "XXX...");
}
x if x < 20 => {
secret.replace_range(2.., "XXXX...");
}
x if x < 30 => {
secret.replace_range(3.., "XXXXX...");
}
_x => {
secret.replace_range(5.., "XXXXXX...");
}
}
}
}
impl Settings<Checked> {
pub fn cleared() -> Settings<Checked> {
Settings {
@ -227,6 +267,7 @@ impl Settings<Checked> {
faceting: Setting::Reset,
pagination: Setting::Reset,
embedders: Setting::Reset,
search_cutoff_ms: Setting::Reset,
_kind: PhantomData,
}
}
@ -249,6 +290,7 @@ impl Settings<Checked> {
faceting,
pagination,
embedders,
search_cutoff_ms,
..
} = self;
@ -269,6 +311,7 @@ impl Settings<Checked> {
faceting,
pagination,
embedders,
search_cutoff_ms,
_kind: PhantomData,
}
}
@ -315,6 +358,7 @@ impl Settings<Unchecked> {
faceting: self.faceting,
pagination: self.pagination,
embedders: self.embedders,
search_cutoff_ms: self.search_cutoff_ms,
_kind: PhantomData,
}
}
@ -347,19 +391,40 @@ pub fn apply_settings_to_builder(
settings: &Settings<Checked>,
builder: &mut milli::update::Settings,
) {
match settings.searchable_attributes {
let Settings {
displayed_attributes,
searchable_attributes,
filterable_attributes,
sortable_attributes,
ranking_rules,
stop_words,
non_separator_tokens,
separator_tokens,
dictionary,
synonyms,
distinct_attribute,
proximity_precision,
typo_tolerance,
faceting,
pagination,
embedders,
search_cutoff_ms,
_kind,
} = settings;
match searchable_attributes {
Setting::Set(ref names) => builder.set_searchable_fields(names.clone()),
Setting::Reset => builder.reset_searchable_fields(),
Setting::NotSet => (),
}
match settings.displayed_attributes {
match displayed_attributes {
Setting::Set(ref names) => builder.set_displayed_fields(names.clone()),
Setting::Reset => builder.reset_displayed_fields(),
Setting::NotSet => (),
}
match settings.filterable_attributes {
match filterable_attributes {
Setting::Set(ref facets) => {
builder.set_filterable_fields(facets.clone().into_iter().collect())
}
@ -367,13 +432,13 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.sortable_attributes {
match sortable_attributes {
Setting::Set(ref fields) => builder.set_sortable_fields(fields.iter().cloned().collect()),
Setting::Reset => builder.reset_sortable_fields(),
Setting::NotSet => (),
}
match settings.ranking_rules {
match ranking_rules {
Setting::Set(ref criteria) => {
builder.set_criteria(criteria.iter().map(|c| c.clone().into()).collect())
}
@ -381,13 +446,13 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.stop_words {
match stop_words {
Setting::Set(ref stop_words) => builder.set_stop_words(stop_words.clone()),
Setting::Reset => builder.reset_stop_words(),
Setting::NotSet => (),
}
match settings.non_separator_tokens {
match non_separator_tokens {
Setting::Set(ref non_separator_tokens) => {
builder.set_non_separator_tokens(non_separator_tokens.clone())
}
@ -395,7 +460,7 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.separator_tokens {
match separator_tokens {
Setting::Set(ref separator_tokens) => {
builder.set_separator_tokens(separator_tokens.clone())
}
@ -403,31 +468,31 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.dictionary {
match dictionary {
Setting::Set(ref dictionary) => builder.set_dictionary(dictionary.clone()),
Setting::Reset => builder.reset_dictionary(),
Setting::NotSet => (),
}
match settings.synonyms {
match synonyms {
Setting::Set(ref synonyms) => builder.set_synonyms(synonyms.clone().into_iter().collect()),
Setting::Reset => builder.reset_synonyms(),
Setting::NotSet => (),
}
match settings.distinct_attribute {
match distinct_attribute {
Setting::Set(ref attr) => builder.set_distinct_field(attr.clone()),
Setting::Reset => builder.reset_distinct_field(),
Setting::NotSet => (),
}
match settings.proximity_precision {
match proximity_precision {
Setting::Set(ref precision) => builder.set_proximity_precision((*precision).into()),
Setting::Reset => builder.reset_proximity_precision(),
Setting::NotSet => (),
}
match settings.typo_tolerance {
match typo_tolerance {
Setting::Set(ref value) => {
match value.enabled {
Setting::Set(val) => builder.set_autorize_typos(val),
@ -482,7 +547,7 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match &settings.faceting {
match faceting {
Setting::Set(FacetingSettings { max_values_per_facet, sort_facet_values_by }) => {
match max_values_per_facet {
Setting::Set(val) => builder.set_max_values_per_facet(*val),
@ -504,7 +569,7 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.pagination {
match pagination {
Setting::Set(ref value) => match value.max_total_hits {
Setting::Set(val) => builder.set_pagination_max_total_hits(val),
Setting::Reset => builder.reset_pagination_max_total_hits(),
@ -514,16 +579,28 @@ pub fn apply_settings_to_builder(
Setting::NotSet => (),
}
match settings.embedders.clone() {
Setting::Set(value) => builder.set_embedder_settings(value),
match embedders {
Setting::Set(value) => builder.set_embedder_settings(value.clone()),
Setting::Reset => builder.reset_embedder_settings(),
Setting::NotSet => (),
}
match search_cutoff_ms {
Setting::Set(cutoff) => builder.set_search_cutoff(*cutoff),
Setting::Reset => builder.reset_search_cutoff(),
Setting::NotSet => (),
}
}
pub enum SecretPolicy {
RevealSecrets,
HideSecrets,
}
pub fn settings(
index: &Index,
rtxn: &crate::heed::RoTxn,
secret_policy: SecretPolicy,
) -> Result<Settings<Checked>, milli::Error> {
let displayed_attributes =
index.displayed_fields(rtxn)?.map(|fields| fields.into_iter().map(String::from).collect());
@ -607,7 +684,9 @@ pub fn settings(
.collect();
let embedders = if embedders.is_empty() { Setting::NotSet } else { Setting::Set(embedders) };
Ok(Settings {
let search_cutoff_ms = index.search_cutoff(rtxn)?;
let mut settings = Settings {
displayed_attributes: match displayed_attributes {
Some(attrs) => Setting::Set(attrs),
None => Setting::Reset,
@ -633,8 +712,18 @@ pub fn settings(
faceting: Setting::Set(faceting),
pagination: Setting::Set(pagination),
embedders,
search_cutoff_ms: match search_cutoff_ms {
Some(cutoff) => Setting::Set(cutoff),
None => Setting::Reset,
},
_kind: PhantomData,
})
};
if let SecretPolicy::HideSecrets = secret_policy {
settings.hide_secrets()
}
Ok(settings)
}
#[derive(Debug, Clone, PartialEq, Eq, Deserr)]
@ -783,6 +872,7 @@ pub(crate) mod test {
faceting: Setting::NotSet,
pagination: Setting::NotSet,
embedders: Setting::NotSet,
search_cutoff_ms: Setting::NotSet,
_kind: PhantomData::<Unchecked>,
};
@ -809,6 +899,7 @@ pub(crate) mod test {
faceting: Setting::NotSet,
pagination: Setting::NotSet,
embedders: Setting::NotSet,
search_cutoff_ms: Setting::NotSet,
_kind: PhantomData::<Unchecked>,
};

3
meilisearch-types/src/task_view.rs
View File

@ -86,7 +86,8 @@ impl From<Details> for DetailsView {
..DetailsView::default()
}
}
Details::SettingsUpdate { settings } => {
Details::SettingsUpdate { mut settings } => {
settings.hide_secrets();
DetailsView { settings: Some(settings), ..DetailsView::default() }
}
Details::IndexInfo { primary_key } => {

12
meilisearch/Cargo.toml
View File

@ -14,18 +14,18 @@ default-run = "meilisearch"
[dependencies]
actix-cors = "0.7.0"
actix-http = { version = "3.5.1", default-features = false, features = [
actix-http = { version = "3.6.0", default-features = false, features = [
"compress-brotli",
"compress-gzip",
"rustls",
"rustls-0_21",
] }
actix-utils = "3.0.1"
actix-web = { version = "4.4.1", default-features = false, features = [
actix-web = { version = "4.5.1", default-features = false, features = [
"macros",
"compress-brotli",
"compress-gzip",
"cookies",
"rustls",
"rustls-0_21",
] }
actix-web-static-files = { git = "https://github.com/kilork/actix-web-static-files.git", rev = "2d3b6160", optional = true }
anyhow = { version = "1.0.79", features = ["backtrace"] }
@ -52,7 +52,7 @@ index-scheduler = { path = "../index-scheduler" }
indexmap = { version = "2.1.0", features = ["serde"] }
is-terminal = "0.4.10"
itertools = "0.11.0"
jsonwebtoken = "8.3.0"
jsonwebtoken = "9.2.0"
lazy_static = "1.4.0"
meilisearch-auth = { path = "../meilisearch-auth" }
meilisearch-types = { path = "../meilisearch-types" }
@ -75,7 +75,7 @@ reqwest = { version = "0.11.23", features = [
"rustls-tls",
"json",
], default-features = false }
rustls = "0.20.8"
rustls = "0.21.6"
rustls-pemfile = "1.0.2"
segment = { version = "0.2.3", optional = true }
serde = { version = "1.0.195", features = ["derive"] }

24
meilisearch/src/analytics/segment_analytics.rs
View File

@ -252,6 +252,7 @@ impl super::Analytics for SegmentAnalytics {
struct Infos {
env: String,
experimental_enable_metrics: bool,
experimental_search_queue_size: usize,
experimental_logs_mode: LogMode,
experimental_replication_parameters: bool,
experimental_enable_logs_route: bool,
@ -293,6 +294,7 @@ impl From<Opt> for Infos {
let Opt {
db_path,
experimental_enable_metrics,
experimental_search_queue_size,
experimental_logs_mode,
experimental_replication_parameters,
experimental_enable_logs_route,
@ -342,6 +344,7 @@ impl From<Opt> for Infos {
Self {
env,
experimental_enable_metrics,
experimental_search_queue_size,
experimental_logs_mode,
experimental_replication_parameters,
experimental_enable_logs_route,
@ -579,6 +582,8 @@ pub struct SearchAggregator {
// requests
total_received: usize,
total_succeeded: usize,
total_degraded: usize,
total_used_negative_operator: usize,
time_spent: BinaryHeap<usize>,
// sort
@ -753,14 +758,22 @@ impl SearchAggregator {
let SearchResult {
hits: _,
query: _,
vector: _,
processing_time_ms,
hits_info: _,
semantic_hit_count: _,
facet_distribution: _,
facet_stats: _,
degraded,
used_negative_operator,
} = result;
self.total_succeeded = self.total_succeeded.saturating_add(1);
if *degraded {
self.total_degraded = self.total_degraded.saturating_add(1);
}
if *used_negative_operator {
self.total_used_negative_operator = self.total_used_negative_operator.saturating_add(1);
}
self.time_spent.push(*processing_time_ms as usize);
}
@ -802,6 +815,8 @@ impl SearchAggregator {
semantic_ratio,
embedder,
hybrid,
total_degraded,
total_used_negative_operator,
} = other;
if self.timestamp.is_none() {
@ -816,6 +831,9 @@ impl SearchAggregator {
// request
self.total_received = self.total_received.saturating_add(total_received);
self.total_succeeded = self.total_succeeded.saturating_add(total_succeeded);
self.total_degraded = self.total_degraded.saturating_add(total_degraded);
self.total_used_negative_operator =
self.total_used_negative_operator.saturating_add(total_used_negative_operator);
self.time_spent.append(time_spent);
// sort
@ -921,6 +939,8 @@ impl SearchAggregator {
semantic_ratio,
embedder,
hybrid,
total_degraded,
total_used_negative_operator,
} = self;
if total_received == 0 {
@ -940,6 +960,8 @@ impl SearchAggregator {
"total_succeeded": total_succeeded,
"total_failed": total_received.saturating_sub(total_succeeded), // just to be sure we never panics
"total_received": total_received,
"total_degraded": total_degraded,
"total_used_negative_operator": total_used_negative_operator,
},
"sort": {
"with_geoPoint": sort_with_geo_point,

6
meilisearch/src/error.rs
View File

@ -29,6 +29,10 @@ pub enum MeilisearchHttpError {
InvalidExpression(&'static [&'static str], Value),
#[error("A {0} payload is missing.")]
MissingPayload(PayloadType),
#[error("Too many search requests running at the same time: {0}. Retry after 10s.")]
TooManySearchRequests(usize),
#[error("Internal error: Search limiter is down.")]
SearchLimiterIsDown,
#[error("The provided payload reached the size limit. The maximum accepted payload size is {}.", Byte::from_bytes(*.0 as u64).get_appropriate_unit(true))]
PayloadTooLarge(usize),
#[error("Two indexes must be given for each swap. The list `[{}]` contains {} indexes.",
@ -69,6 +73,8 @@ impl ErrorCode for MeilisearchHttpError {
MeilisearchHttpError::EmptyFilter => Code::InvalidDocumentFilter,
MeilisearchHttpError::InvalidExpression(_, _) => Code::InvalidSearchFilter,
MeilisearchHttpError::PayloadTooLarge(_) => Code::PayloadTooLarge,
MeilisearchHttpError::TooManySearchRequests(_) => Code::TooManySearchRequests,
MeilisearchHttpError::SearchLimiterIsDown => Code::Internal,
MeilisearchHttpError::SwapIndexPayloadWrongLength(_) => Code::InvalidSwapIndexes,
MeilisearchHttpError::IndexUid(e) => e.error_code(),
MeilisearchHttpError::SerdeJson(_) => Code::Internal,

10
meilisearch/src/lib.rs
View File

@ -9,12 +9,14 @@ pub mod middleware;
pub mod option;
pub mod routes;
pub mod search;
pub mod search_queue;
use std::fs::File;
use std::io::{BufReader, BufWriter};
use std::num::NonZeroUsize;
use std::path::Path;
use std::sync::Arc;
use std::thread;
use std::thread::{self, available_parallelism};
use std::time::Duration;
use actix_cors::Cors;
@ -38,6 +40,7 @@ use meilisearch_types::versioning::{check_version_file, create_version_file};
use meilisearch_types::{compression, milli, VERSION_FILE_NAME};
pub use option::Opt;
use option::ScheduleSnapshot;
use search_queue::SearchQueue;
use tracing::{error, info_span};
use tracing_subscriber::filter::Targets;
@ -469,10 +472,15 @@ pub fn configure_data(
(logs_route, logs_stderr): (LogRouteHandle, LogStderrHandle),
analytics: Arc<dyn Analytics>,
) {
let search_queue = SearchQueue::new(
opt.experimental_search_queue_size,
available_parallelism().unwrap_or(NonZeroUsize::new(2).unwrap()),
);
let http_payload_size_limit = opt.http_payload_size_limit.get_bytes() as usize;
config
.app_data(index_scheduler)
.app_data(auth)
.app_data(web::Data::new(search_queue))
.app_data(web::Data::from(analytics))
.app_data(web::Data::new(logs_route))
.app_data(web::Data::new(logs_stderr))

2
meilisearch/src/main.rs
View File

@ -151,7 +151,7 @@ async fn run_http(
.keep_alive(KeepAlive::Os);
if let Some(config) = opt_clone.get_ssl_config()? {
http_server.bind_rustls(opt_clone.http_addr, config)?.run().await?;
http_server.bind_rustls_021(opt_clone.http_addr, config)?.run().await?;
} else {
http_server.bind(&opt_clone.http_addr)?.run().await?;
}

21
meilisearch/src/metrics.rs
View File

@ -4,24 +4,17 @@ use prometheus::{
register_int_gauge_vec, HistogramVec, IntCounterVec, IntGauge, IntGaugeVec,
};
/// Create evenly distributed buckets
fn create_buckets() -> [f64; 29] {
(0..10)
.chain((10..100).step_by(10))
.chain((100..=1000).step_by(100))
.map(|i| i as f64 / 1000.)
.collect::<Vec<_>>()
.try_into()
.unwrap()
}
lazy_static! {
pub static ref MEILISEARCH_HTTP_RESPONSE_TIME_CUSTOM_BUCKETS: [f64; 29] = create_buckets();
pub static ref MEILISEARCH_HTTP_REQUESTS_TOTAL: IntCounterVec = register_int_counter_vec!(
opts!("meilisearch_http_requests_total", "Meilisearch HTTP requests total"),
&["method", "path"]
&["method", "path", "status"]
)
.expect("Can't create a metric");
pub static ref MEILISEARCH_DEGRADED_SEARCH_REQUESTS: IntGauge = register_int_gauge!(opts!(
"meilisearch_degraded_search_requests",
"Meilisearch number of degraded search requests"
))
.expect("Can't create a metric");
pub static ref MEILISEARCH_DB_SIZE_BYTES: IntGauge =
register_int_gauge!(opts!("meilisearch_db_size_bytes", "Meilisearch DB Size In Bytes"))
.expect("Can't create a metric");
@ -42,7 +35,7 @@ lazy_static! {
"meilisearch_http_response_time_seconds",
"Meilisearch HTTP response times",
&["method", "path"],
MEILISEARCH_HTTP_RESPONSE_TIME_CUSTOM_BUCKETS.to_vec()
vec![0.005, 0.01, 0.025, 0.05, 0.075, 0.1, 0.25, 0.5, 0.75, 1.0, 2.5, 5.0, 7.5, 10.0]
)
.expect("Can't create a metric");
pub static ref MEILISEARCH_NB_TASKS: IntGaugeVec = register_int_gauge_vec!(

11
meilisearch/src/middleware.rs
View File

@ -65,9 +65,6 @@ where
.with_label_values(&[&request_method, request_path])
.start_timer(),
);
crate::metrics::MEILISEARCH_HTTP_REQUESTS_TOTAL
.with_label_values(&[&request_method, request_path])
.inc();
}
};
@ -76,6 +73,14 @@ where
Box::pin(async move {
let res = fut.await?;
crate::metrics::MEILISEARCH_HTTP_REQUESTS_TOTAL
.with_label_values(&[
res.request().method().as_str(),
res.request().path(),
res.status().as_str(),
])
.inc();
if let Some(histogram_timer) = histogram_timer {
histogram_timer.observe_duration();
};

19
meilisearch/src/option.rs
View File

@ -54,6 +54,7 @@ const MEILI_EXPERIMENTAL_LOGS_MODE: &str = "MEILI_EXPERIMENTAL_LOGS_MODE";
const MEILI_EXPERIMENTAL_REPLICATION_PARAMETERS: &str = "MEILI_EXPERIMENTAL_REPLICATION_PARAMETERS";
const MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE: &str = "MEILI_EXPERIMENTAL_ENABLE_LOGS_ROUTE";
const MEILI_EXPERIMENTAL_ENABLE_METRICS: &str = "MEILI_EXPERIMENTAL_ENABLE_METRICS";
const MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE: &str = "MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE";
const MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE: &str =
"MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE";
const MEILI_EXPERIMENTAL_MAX_NUMBER_OF_BATCHED_TASKS: &str =
@ -344,6 +345,15 @@ pub struct Opt {
#[serde(default)]
pub experimental_enable_metrics: bool,
/// Experimental search queue size. For more information, see: <https://github.com/orgs/meilisearch/discussions/729>
///
/// Lets you customize the size of the search queue. Meilisearch processes your search requests as fast as possible but once the
/// queue is full it starts returning HTTP 503, Service Unavailable.
/// The default value is 1000.
#[clap(long, env = MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE, default_value_t = 1000)]
#[serde(default)]
pub experimental_search_queue_size: usize,
/// Experimental logs mode feature. For more information, see: <https://github.com/orgs/meilisearch/discussions/723>
///
/// Change the mode of the logs on the console.
@ -473,6 +483,7 @@ impl Opt {
#[cfg(feature = "analytics")]
no_analytics,
experimental_enable_metrics,
experimental_search_queue_size,
experimental_logs_mode,
experimental_enable_logs_route,
experimental_replication_parameters,
@ -532,6 +543,10 @@ impl Opt {
MEILI_EXPERIMENTAL_ENABLE_METRICS,
experimental_enable_metrics.to_string(),
);
export_to_env_if_not_present(
MEILI_EXPERIMENTAL_SEARCH_QUEUE_SIZE,
experimental_search_queue_size.to_string(),
);
export_to_env_if_not_present(
MEILI_EXPERIMENTAL_LOGS_MODE,
experimental_logs_mode.to_string(),
@ -564,11 +579,11 @@ impl Opt {
}
if self.ssl_require_auth {
let verifier = AllowAnyAuthenticatedClient::new(client_auth_roots);
config.with_client_cert_verifier(verifier)
config.with_client_cert_verifier(Arc::from(verifier))
} else {
let verifier =
AllowAnyAnonymousOrAuthenticatedClient::new(client_auth_roots);
config.with_client_cert_verifier(verifier)
config.with_client_cert_verifier(Arc::from(verifier))
}
}
None => config.with_no_client_auth(),

7
meilisearch/src/routes/indexes/facet_search.rs
View File

@ -12,11 +12,13 @@ use tracing::debug;
use crate::analytics::{Analytics, FacetSearchAggregator};
use crate::extractors::authentication::policies::*;
use crate::extractors::authentication::GuardedData;
use crate::routes::indexes::search::search_kind;
use crate::search::{
add_search_rules, perform_facet_search, HybridQuery, MatchingStrategy, SearchQuery,
DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG,
DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET,
};
use crate::search_queue::SearchQueue;
pub fn configure(cfg: &mut web::ServiceConfig) {
cfg.service(web::resource("").route(web::post().to(search)));
@ -48,6 +50,7 @@ pub struct FacetSearchQuery {
pub async fn search(
index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
search_queue: Data<SearchQueue>,
index_uid: web::Path<String>,
params: AwebJson<FacetSearchQuery, DeserrJsonError>,
req: HttpRequest,
@ -71,8 +74,10 @@ pub async fn search(
let index = index_scheduler.index(&index_uid)?;
let features = index_scheduler.features();
let search_kind = search_kind(&search_query, &index_scheduler, &index, features)?;
let _permit = search_queue.try_get_search_permit().await?;
let search_result = tokio::task::spawn_blocking(move || {
perform_facet_search(&index, search_query, facet_query, facet_name, features)
perform_facet_search(&index, search_query, facet_query, facet_name, search_kind)
})
.await?;

138
meilisearch/src/routes/indexes/search.rs
View File

@ -1,27 +1,29 @@
use actix_web::web::Data;
use actix_web::{web, HttpRequest, HttpResponse};
use deserr::actix_web::{AwebJson, AwebQueryParameter};
use index_scheduler::IndexScheduler;
use index_scheduler::{IndexScheduler, RoFeatures};
use meilisearch_types::deserr::query_params::Param;
use meilisearch_types::deserr::{DeserrJsonError, DeserrQueryParamError};
use meilisearch_types::error::deserr_codes::*;
use meilisearch_types::error::ResponseError;
use meilisearch_types::index_uid::IndexUid;
use meilisearch_types::milli;
use meilisearch_types::milli::vector::DistributionShift;
use meilisearch_types::serde_cs::vec::CS;
use serde_json::Value;
use tracing::{debug, warn};
use tracing::debug;
use crate::analytics::{Analytics, SearchAggregator};
use crate::error::MeilisearchHttpError;
use crate::extractors::authentication::policies::*;
use crate::extractors::authentication::GuardedData;
use crate::extractors::sequential_extractor::SeqHandler;
use crate::metrics::MEILISEARCH_DEGRADED_SEARCH_REQUESTS;
use crate::search::{
add_search_rules, perform_search, HybridQuery, MatchingStrategy, SearchQuery, SemanticRatio,
DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG,
add_search_rules, perform_search, HybridQuery, MatchingStrategy, SearchKind, SearchQuery,
SemanticRatio, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG,
DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET, DEFAULT_SEMANTIC_RATIO,
};
use crate::search_queue::SearchQueue;
pub fn configure(cfg: &mut web::ServiceConfig) {
cfg.service(
@ -181,6 +183,7 @@ fn fix_sort_query_parameters(sort_query: &str) -> Vec<String> {
pub async fn search_with_url_query(
index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
search_queue: web::Data<SearchQueue>,
index_uid: web::Path<String>,
params: AwebQueryParameter<SearchQueryGet, DeserrQueryParamError>,
req: HttpRequest,
@ -201,11 +204,11 @@ pub async fn search_with_url_query(
let index = index_scheduler.index(&index_uid)?;
let features = index_scheduler.features();
let distribution = embed(&mut query, index_scheduler.get_ref(), &index).await?;
let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)?;
let _permit = search_queue.try_get_search_permit().await?;
let search_result =
tokio::task::spawn_blocking(move || perform_search(&index, query, features, distribution))
.await?;
tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind)).await?;
if let Ok(ref search_result) = search_result {
aggregate.succeed(search_result);
}
@ -219,6 +222,7 @@ pub async fn search_with_url_query(
pub async fn search_with_post(
index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
search_queue: web::Data<SearchQueue>,
index_uid: web::Path<String>,
params: AwebJson<SearchQuery, DeserrJsonError>,
req: HttpRequest,
@ -240,13 +244,16 @@ pub async fn search_with_post(
let features = index_scheduler.features();
let distribution = embed(&mut query, index_scheduler.get_ref(), &index).await?;
let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)?;
let _permit = search_queue.try_get_search_permit().await?;
let search_result =
tokio::task::spawn_blocking(move || perform_search(&index, query, features, distribution))
.await?;
tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind)).await?;
if let Ok(ref search_result) = search_result {
aggregate.succeed(search_result);
if search_result.degraded {
MEILISEARCH_DEGRADED_SEARCH_REQUESTS.inc();
}
}
analytics.post_search(aggregate);
@ -256,77 +263,58 @@ pub async fn search_with_post(
Ok(HttpResponse::Ok().json(search_result))
}
pub async fn embed(
query: &mut SearchQuery,
pub fn search_kind(
query: &SearchQuery,
index_scheduler: &IndexScheduler,
index: &milli::Index,
) -> Result<Option<DistributionShift>, ResponseError> {
match (&query.hybrid, &query.vector, &query.q) {
(Some(HybridQuery { semantic_ratio: _, embedder }), None, Some(q))
if !q.trim().is_empty() =>
{
let embedder_configs = index.embedding_configs(&index.read_txn()?)?;
let embedders = index_scheduler.embedders(embedder_configs)?;
features: RoFeatures,
) -> Result<SearchKind, ResponseError> {
if query.vector.is_some() {
features.check_vector("Passing `vector` as a query parameter")?;
}
let embedder = if let Some(embedder_name) = embedder {
embedders.get(embedder_name)
} else {
embedders.get_default()
};
if query.hybrid.is_some() {
features.check_vector("Passing `hybrid` as a query parameter")?;
}
let embedder = embedder
.ok_or(milli::UserError::InvalidEmbedder("default".to_owned()))
.map_err(milli::Error::from)?
.0;
let distribution = embedder.distribution();
let embeddings = embedder
.embed(vec![q.to_owned()])
.await
.map_err(milli::vector::Error::from)
.map_err(milli::Error::from)?
.pop()
.expect("No vector returned from embedding");
if embeddings.iter().nth(1).is_some() {
warn!("Ignoring embeddings past the first one in long search query");
query.vector = Some(embeddings.iter().next().unwrap().to_vec());
} else {
query.vector = Some(embeddings.into_inner());
}
Ok(distribution)
// regardless of anything, always do a keyword search when we don't have a vector and the query is whitespace or missing
if query.vector.is_none() {
match &query.q {
Some(q) if q.trim().is_empty() => return Ok(SearchKind::KeywordOnly),
None => return Ok(SearchKind::KeywordOnly),
_ => {}
}
(Some(hybrid), vector, _) => {
let embedder_configs = index.embedding_configs(&index.read_txn()?)?;
let embedders = index_scheduler.embedders(embedder_configs)?;
}
let embedder = if let Some(embedder_name) = &hybrid.embedder {
embedders.get(embedder_name)
} else {
embedders.get_default()
};
let embedder = embedder
.ok_or(milli::UserError::InvalidEmbedder("default".to_owned()))
.map_err(milli::Error::from)?
.0;
if let Some(vector) = vector {
if vector.len() != embedder.dimensions() {
return Err(meilisearch_types::milli::Error::UserError(
meilisearch_types::milli::UserError::InvalidVectorDimensions {
expected: embedder.dimensions(),
found: vector.len(),
},
)
.into());
}
}
Ok(embedder.distribution())
match &query.hybrid {
Some(HybridQuery { semantic_ratio, embedder }) if **semantic_ratio == 1.0 => {
Ok(SearchKind::semantic(
index_scheduler,
index,
embedder.as_deref(),
query.vector.as_ref().map(Vec::len),
)?)
}
_ => Ok(None),
Some(HybridQuery { semantic_ratio, embedder: _ }) if **semantic_ratio == 0.0 => {
Ok(SearchKind::KeywordOnly)
}
Some(HybridQuery { semantic_ratio, embedder }) => Ok(SearchKind::hybrid(
index_scheduler,
index,
embedder.as_deref(),
**semantic_ratio,
query.vector.as_ref().map(Vec::len),
)?),
None => match (query.q.as_deref(), query.vector.as_deref()) {
(_query, None) => Ok(SearchKind::KeywordOnly),
(None, Some(_vector)) => Ok(SearchKind::semantic(
index_scheduler,
index,
None,
query.vector.as_ref().map(Vec::len),
)?),
(Some(_), Some(_)) => Err(MeilisearchHttpError::MissingSearchHybrid.into()),
},
}
}

33
meilisearch/src/routes/indexes/settings.rs
View File

@ -7,7 +7,7 @@ use meilisearch_types::error::ResponseError;
use meilisearch_types::facet_values_sort::FacetValuesSort;
use meilisearch_types::index_uid::IndexUid;
use meilisearch_types::milli::update::Setting;
use meilisearch_types::settings::{settings, RankingRuleView, Settings, Unchecked};
use meilisearch_types::settings::{settings, RankingRuleView, SecretPolicy, Settings, Unchecked};
use meilisearch_types::tasks::KindWithContent;
use serde_json::json;
use tracing::debug;
@ -134,7 +134,7 @@ macro_rules! make_setting_route {
let index = index_scheduler.index(&index_uid)?;
let rtxn = index.read_txn()?;
let settings = settings(&index, &rtxn)?;
let settings = settings(&index, &rtxn, meilisearch_types::settings::SecretPolicy::HideSecrets)?;
debug!(returns = ?settings, "Update settings");
let mut json = serde_json::json!(&settings);
@ -604,6 +604,8 @@ fn embedder_analytics(
EmbedderSource::OpenAi => sources.insert("openAi"),
EmbedderSource::HuggingFace => sources.insert("huggingFace"),
EmbedderSource::UserProvided => sources.insert("userProvided"),
EmbedderSource::Ollama => sources.insert("ollama"),
EmbedderSource::Rest => sources.insert("rest"),
};
}
};
@ -623,6 +625,25 @@ fn embedder_analytics(
)
}
make_setting_route!(
"/search-cutoff-ms",
put,
u64,
meilisearch_types::deserr::DeserrJsonError<
meilisearch_types::error::deserr_codes::InvalidSettingsSearchCutoffMs,
>,
search_cutoff_ms,
"searchCutoffMs",
analytics,
|setting: &Option<u64>, req: &HttpRequest| {
analytics.publish(
"Search Cutoff Updated".to_string(),
serde_json::json!({"search_cutoff_ms": setting }),
Some(req),
);
}
);
macro_rules! generate_configure {
($($mod:ident),*) => {
pub fn configure(cfg: &mut web::ServiceConfig) {
@ -653,7 +674,8 @@ generate_configure!(
typo_tolerance,
pagination,
faceting,
embedders
embedders,
search_cutoff_ms
);
pub async fn update_all(
@ -764,7 +786,8 @@ pub async fn update_all(
"synonyms": {
"total": new_settings.synonyms.as_ref().set().map(|synonyms| synonyms.len()),
},
"embedders": crate::routes::indexes::settings::embedder_analytics(new_settings.embedders.as_ref().set())
"embedders": crate::routes::indexes::settings::embedder_analytics(new_settings.embedders.as_ref().set()),
"search_cutoff_ms": new_settings.search_cutoff_ms.as_ref().set(),
}),
Some(&req),
);
@ -796,7 +819,7 @@ pub async fn get_all(
let index = index_scheduler.index(&index_uid)?;
let rtxn = index.read_txn()?;
let new_settings = settings(&index, &rtxn)?;
let new_settings = settings(&index, &rtxn, SecretPolicy::HideSecrets)?;
debug!(returns = ?new_settings, "Get all settings");
Ok(HttpResponse::Ok().json(new_settings))
}

3
meilisearch/src/routes/mod.rs
View File

@ -15,6 +15,7 @@ use tracing::debug;
use crate::analytics::Analytics;
use crate::extractors::authentication::policies::*;
use crate::extractors::authentication::GuardedData;
use crate::search_queue::SearchQueue;
use crate::Opt;
const PAGINATION_DEFAULT_LIMIT: usize = 20;
@ -385,10 +386,12 @@ pub async fn get_health(
req: HttpRequest,
index_scheduler: Data<IndexScheduler>,
auth_controller: Data<AuthController>,
search_queue: Data<SearchQueue>,
analytics: web::Data<dyn Analytics>,
) -> Result<HttpResponse, ResponseError> {
analytics.health_seen(&req);
search_queue.health().unwrap();
index_scheduler.health().unwrap();
auth_controller.health().unwrap();

20
meilisearch/src/routes/multi_search.rs
View File

@ -13,10 +13,11 @@ use crate::analytics::{Analytics, MultiSearchAggregator};
use crate::extractors::authentication::policies::ActionPolicy;
use crate::extractors::authentication::{AuthenticationError, GuardedData};
use crate::extractors::sequential_extractor::SeqHandler;
use crate::routes::indexes::search::embed;
use crate::routes::indexes::search::search_kind;
use crate::search::{
add_search_rules, perform_search, SearchQueryWithIndex, SearchResultWithIndex,
};
use crate::search_queue::SearchQueue;
pub fn configure(cfg: &mut web::ServiceConfig) {
cfg.service(web::resource("").route(web::post().to(SeqHandler(multi_search_with_post))));
@ -35,6 +36,7 @@ pub struct SearchQueries {
pub async fn multi_search_with_post(
index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
search_queue: Data<SearchQueue>,
params: AwebJson<SearchQueries, DeserrJsonError>,
req: HttpRequest,
analytics: web::Data<dyn Analytics>,
@ -44,6 +46,10 @@ pub async fn multi_search_with_post(
let mut multi_aggregate = MultiSearchAggregator::from_queries(&queries, &req);
let features = index_scheduler.features();
// Since we don't want to process half of the search requests and then get a permit refused
// we're going to get one permit for the whole duration of the multi-search request.
let _permit = search_queue.try_get_search_permit().await?;
// Explicitly expect a `(ResponseError, usize)` for the error type rather than `ResponseError` only,
// so that `?` doesn't work if it doesn't use `with_index`, ensuring that it is not forgotten in case of code
// changes.
@ -75,15 +81,13 @@ pub async fn multi_search_with_post(
})
.with_index(query_index)?;
let distribution = embed(&mut query, index_scheduler.get_ref(), &index)
.await
let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)
.with_index(query_index)?;
let search_result = tokio::task::spawn_blocking(move || {
perform_search(&index, query, features, distribution)
})
.await
.with_index(query_index)?;
let search_result =
tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind))
.await
.with_index(query_index)?;
search_results.push(SearchResultWithIndex {
index_uid: index_uid.into_inner(),

238
meilisearch/src/search.rs
View File

@ -1,20 +1,21 @@
use std::cmp::min;
use std::collections::{BTreeMap, BTreeSet, HashSet};
use std::str::FromStr;
use std::time::Instant;
use std::sync::Arc;
use std::time::{Duration, Instant};
use deserr::Deserr;
use either::Either;
use index_scheduler::RoFeatures;
use indexmap::IndexMap;
use meilisearch_auth::IndexSearchRules;
use meilisearch_types::deserr::DeserrJsonError;
use meilisearch_types::error::deserr_codes::*;
use meilisearch_types::error::ResponseError;
use meilisearch_types::heed::RoTxn;
use meilisearch_types::index_uid::IndexUid;
use meilisearch_types::milli::score_details::{self, ScoreDetails, ScoringStrategy};
use meilisearch_types::milli::vector::DistributionShift;
use meilisearch_types::milli::{FacetValueHit, OrderBy, SearchForFacetValues};
use meilisearch_types::milli::score_details::{ScoreDetails, ScoringStrategy};
use meilisearch_types::milli::vector::Embedder;
use meilisearch_types::milli::{FacetValueHit, OrderBy, SearchForFacetValues, TimeBudget};
use meilisearch_types::settings::DEFAULT_PAGINATION_MAX_TOTAL_HITS;
use meilisearch_types::{milli, Document};
use milli::tokenizer::TokenizerBuilder;
@ -90,13 +91,75 @@ pub struct SearchQuery {
#[derive(Debug, Clone, Default, PartialEq, Deserr)]
#[deserr(error = DeserrJsonError<InvalidHybridQuery>, rename_all = camelCase, deny_unknown_fields)]
pub struct HybridQuery {
/// TODO validate that sementic ratio is between 0.0 and 1,0
#[deserr(default, error = DeserrJsonError<InvalidSearchSemanticRatio>, default)]
pub semantic_ratio: SemanticRatio,
#[deserr(default, error = DeserrJsonError<InvalidEmbedder>, default)]
pub embedder: Option<String>,
}
pub enum SearchKind {
KeywordOnly,
SemanticOnly { embedder_name: String, embedder: Arc<Embedder> },
Hybrid { embedder_name: String, embedder: Arc<Embedder>, semantic_ratio: f32 },
}
impl SearchKind {
pub(crate) fn semantic(
index_scheduler: &index_scheduler::IndexScheduler,
index: &Index,
embedder_name: Option<&str>,
vector_len: Option<usize>,
) -> Result<Self, ResponseError> {
let (embedder_name, embedder) =
Self::embedder(index_scheduler, index, embedder_name, vector_len)?;
Ok(Self::SemanticOnly { embedder_name, embedder })
}
pub(crate) fn hybrid(
index_scheduler: &index_scheduler::IndexScheduler,
index: &Index,
embedder_name: Option<&str>,
semantic_ratio: f32,
vector_len: Option<usize>,
) -> Result<Self, ResponseError> {
let (embedder_name, embedder) =
Self::embedder(index_scheduler, index, embedder_name, vector_len)?;
Ok(Self::Hybrid { embedder_name, embedder, semantic_ratio })
}
fn embedder(
index_scheduler: &index_scheduler::IndexScheduler,
index: &Index,
embedder_name: Option<&str>,
vector_len: Option<usize>,
) -> Result<(String, Arc<Embedder>), ResponseError> {
let embedder_configs = index.embedding_configs(&index.read_txn()?)?;
let embedders = index_scheduler.embedders(embedder_configs)?;
let embedder_name = embedder_name.unwrap_or_else(|| embedders.get_default_embedder_name());
let embedder = embedders.get(embedder_name);
let embedder = embedder
.ok_or(milli::UserError::InvalidEmbedder(embedder_name.to_owned()))
.map_err(milli::Error::from)?
.0;
if let Some(vector_len) = vector_len {
if vector_len != embedder.dimensions() {
return Err(meilisearch_types::milli::Error::UserError(
meilisearch_types::milli::UserError::InvalidVectorDimensions {
expected: embedder.dimensions(),
found: vector_len,
},
)
.into());
}
}
Ok((embedder_name.to_owned(), embedder))
}
}
#[derive(Debug, Clone, Copy, PartialEq, Deserr)]
#[deserr(try_from(f32) = TryFrom::try_from -> InvalidSearchSemanticRatio)]
pub struct SemanticRatio(f32);
@ -305,8 +368,6 @@ pub struct SearchHit {
pub ranking_score: Option<f64>,
#[serde(rename = "_rankingScoreDetails", skip_serializing_if = "Option::is_none")]
pub ranking_score_details: Option<serde_json::Map<String, serde_json::Value>>,
#[serde(rename = "_semanticScore", skip_serializing_if = "Option::is_none")]
pub semantic_score: Option<f32>,
}
#[derive(Serialize, Debug, Clone, PartialEq)]
@ -314,8 +375,6 @@ pub struct SearchHit {
pub struct SearchResult {
pub hits: Vec<SearchHit>,
pub query: String,
#[serde(skip_serializing_if = "Option::is_none")]
pub vector: Option<Vec<f32>>,
pub processing_time_ms: u128,
#[serde(flatten)]
pub hits_info: HitsInfo,
@ -323,6 +382,15 @@ pub struct SearchResult {
pub facet_distribution: Option<BTreeMap<String, IndexMap<String, u64>>>,
#[serde(skip_serializing_if = "Option::is_none")]
pub facet_stats: Option<BTreeMap<String, FacetStats>>,
#[serde(skip_serializing_if = "Option::is_none")]
pub semantic_hit_count: Option<u32>,
// These fields are only used for analytics purposes
#[serde(skip)]
pub degraded: bool,
#[serde(skip)]
pub used_negative_operator: bool,
}
#[derive(Serialize, Debug, Clone, PartialEq)]
@ -380,45 +448,36 @@ fn prepare_search<'t>(
index: &'t Index,
rtxn: &'t RoTxn,
query: &'t SearchQuery,
features: RoFeatures,
distribution: Option<DistributionShift>,
search_kind: &SearchKind,
time_budget: TimeBudget,
) -> Result<(milli::Search<'t>, bool, usize, usize), MeilisearchHttpError> {
let mut search = index.search(rtxn);
search.time_budget(time_budget);
if query.vector.is_some() {
features.check_vector("Passing `vector` as a query parameter")?;
}
if query.hybrid.is_some() {
features.check_vector("Passing `hybrid` as a query parameter")?;
}
if query.hybrid.is_none() && query.q.is_some() && query.vector.is_some() {
return Err(MeilisearchHttpError::MissingSearchHybrid);
}
search.distribution_shift(distribution);
if let Some(ref vector) = query.vector {
match &query.hybrid {
// If semantic ratio is 0.0, only the query search will impact the search results,
// skip the vector
Some(hybrid) if *hybrid.semantic_ratio == 0.0 => (),
_otherwise => {
search.vector(vector.clone());
}
}
}
if let Some(ref q) = query.q {
match &query.hybrid {
// If semantic ratio is 1.0, only the vector search will impact the search results,
// skip the query
Some(hybrid) if *hybrid.semantic_ratio == 1.0 => (),
_otherwise => {
match search_kind {
SearchKind::KeywordOnly => {
if let Some(q) = &query.q {
search.query(q);
}
}
SearchKind::SemanticOnly { embedder_name, embedder } => {
let vector = match query.vector.clone() {
Some(vector) => vector,
None => embedder
.embed_one(query.q.clone().unwrap())
.map_err(milli::vector::Error::from)
.map_err(milli::Error::from)?,
};
search.semantic(embedder_name.clone(), embedder.clone(), Some(vector));
}
SearchKind::Hybrid { embedder_name, embedder, semantic_ratio: _ } => {
if let Some(q) = &query.q {
search.query(q);
}
// will be embedded in hybrid search if necessary
search.semantic(embedder_name.clone(), embedder.clone(), query.vector.clone());
}
}
if let Some(ref searchable) = query.attributes_to_search_on {
@ -441,10 +500,6 @@ fn prepare_search<'t>(
ScoringStrategy::Skip
});
if let Some(HybridQuery { embedder: Some(embedder), .. }) = &query.hybrid {
search.embedder_name(embedder);
}
// compute the offset on the limit depending on the pagination mode.
let (offset, limit) = if is_finite_pagination {
let limit = query.hits_per_page.unwrap_or_else(DEFAULT_SEARCH_LIMIT);
@ -487,23 +542,37 @@ fn prepare_search<'t>(
pub fn perform_search(
index: &Index,
query: SearchQuery,
features: RoFeatures,
distribution: Option<DistributionShift>,
search_kind: SearchKind,
) -> Result<SearchResult, MeilisearchHttpError> {
let before_search = Instant::now();
let rtxn = index.read_txn()?;
let time_budget = match index.search_cutoff(&rtxn)? {
Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
None => TimeBudget::default(),
};
let (search, is_finite_pagination, max_total_hits, offset) =
prepare_search(index, &rtxn, &query, features, distribution)?;
prepare_search(index, &rtxn, &query, &search_kind, time_budget)?;
let milli::SearchResult { documents_ids, matching_words, candidates, document_scores, .. } =
match &query.hybrid {
Some(hybrid) => match *hybrid.semantic_ratio {
ratio if ratio == 0.0 || ratio == 1.0 => search.execute()?,
ratio => search.execute_hybrid(ratio)?,
},
None => search.execute()?,
};
let (
milli::SearchResult {
documents_ids,
matching_words,
candidates,
document_scores,
degraded,
used_negative_operator,
},
semantic_hit_count,
) = match &search_kind {
SearchKind::KeywordOnly => (search.execute()?, None),
SearchKind::SemanticOnly { .. } => {
let results = search.execute()?;
let semantic_hit_count = results.document_scores.len() as u32;
(results, Some(semantic_hit_count))
}
SearchKind::Hybrid { semantic_ratio, .. } => search.execute_hybrid(*semantic_ratio)?,
};
let fields_ids_map = index.fields_ids_map(&rtxn).unwrap();
@ -530,7 +599,7 @@ pub fn perform_search(
// The attributes to retrieve are the ones explicitly marked as to retrieve (all by default),
// but these attributes must be also be present
// - in the fields_ids_map
// - in the the displayed attributes
// - in the displayed attributes
let to_retrieve_ids: BTreeSet<_> = query
.attributes_to_retrieve
.as_ref()
@ -612,18 +681,6 @@ pub fn perform_search(
insert_geo_distance(sort, &mut document);
}
let mut semantic_score = None;
for details in &score {
if let ScoreDetails::Vector(score_details::Vector {
target_vector: _,
value_similarity: Some((_matching_vector, similarity)),
}) = details
{
semantic_score = Some(*similarity);
break;
}
}
let ranking_score =
query.show_ranking_score.then(|| ScoreDetails::global_score(score.iter()));
let ranking_score_details =
@ -635,7 +692,6 @@ pub fn perform_search(
matches_position,
ranking_score_details,
ranking_score,
semantic_score,
};
documents.push(hit);
}
@ -671,27 +727,16 @@ pub fn perform_search(
let sort_facet_values_by =
index.sort_facet_values_by(&rtxn).map_err(milli::Error::from)?;
let default_sort_facet_values_by =
sort_facet_values_by.get("*").copied().unwrap_or_default();
if fields.iter().all(|f| f != "*") {
let fields: Vec<_> = fields
.iter()
.map(|n| {
(
n,
sort_facet_values_by
.get(n)
.copied()
.unwrap_or(default_sort_facet_values_by),
)
})
.collect();
let fields: Vec<_> =
fields.iter().map(|n| (n, sort_facet_values_by.get(n))).collect();
facet_distribution.facets(fields);
}
let distribution = facet_distribution
.candidates(candidates)
.default_order_by(default_sort_facet_values_by)
.default_order_by(sort_facet_values_by.get("*"))
.execute()?;
let stats = facet_distribution.compute_stats()?;
(Some(distribution), Some(stats))
@ -707,10 +752,12 @@ pub fn perform_search(
hits: documents,
hits_info,
query: query.q.unwrap_or_default(),
vector: query.vector,
processing_time_ms: before_search.elapsed().as_millis(),
facet_distribution,
facet_stats,
degraded,
used_negative_operator,
semantic_hit_count,
};
Ok(result)
}
@ -720,14 +767,21 @@ pub fn perform_facet_search(
search_query: SearchQuery,
facet_query: Option<String>,
facet_name: String,
features: RoFeatures,
search_kind: SearchKind,
) -> Result<FacetSearchResult, MeilisearchHttpError> {
let before_search = Instant::now();
let rtxn = index.read_txn()?;
let time_budget = match index.search_cutoff(&rtxn)? {
Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
None => TimeBudget::default(),
};
let (search, _, _, _) = prepare_search(index, &rtxn, &search_query, features, None)?;
let mut facet_search =
SearchForFacetValues::new(facet_name, search, search_query.hybrid.is_some());
let (search, _, _, _) = prepare_search(index, &rtxn, &search_query, &search_kind, time_budget)?;
let mut facet_search = SearchForFacetValues::new(
facet_name,
search,
matches!(search_kind, SearchKind::Hybrid { .. }),
);
if let Some(facet_query) = &facet_query {
facet_search.query(facet_query);
}

130
meilisearch/src/search_queue.rs Normal file
View File

@ -0,0 +1,130 @@
//! This file implements a queue of searches to process and the ability to control how many searches can be run in parallel.
//! We need this because we don't want to process more search requests than we have cores.
//! That slows down everything and consumes RAM for no reason.
//! The steps to do a search are to get the `SearchQueue` data structure and try to get a search permit.
//! This can fail if the queue is full, and we need to drop your search request to register a new one.
//!
//! ### How to do a search request
//!
//! In order to do a search request you should try to get a search permit.
//! Retrieve the `SearchQueue` structure from actix-web (`search_queue: Data<SearchQueue>`)
//! and right before processing the search, calls the `SearchQueue::try_get_search_permit` method: `search_queue.try_get_search_permit().await?;`
//!
//! What is going to happen at this point is that you're going to send a oneshot::Sender over an async mpsc channel.
//! Then, the queue/scheduler is going to either:
//! - Drop your oneshot channel => that means there are too many searches going on, and yours won't be executed.
//! You should exit and free all the RAM you use ASAP.
//! - Sends you a Permit => that will unlock the method, and you will be able to process your search.
//! And should drop the Permit only once you have freed all the RAM consumed by the method.
use std::num::NonZeroUsize;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use tokio::sync::{mpsc, oneshot};
use crate::error::MeilisearchHttpError;
#[derive(Debug)]
pub struct SearchQueue {
sender: mpsc::Sender<oneshot::Sender<Permit>>,
capacity: usize,
}
/// You should only run search requests while holding this permit.
/// Once it's dropped, a new search request will be able to process.
#[derive(Debug)]
pub struct Permit {
sender: mpsc::Sender<()>,
}
impl Drop for Permit {
fn drop(&mut self) {
// if the channel is closed then the whole instance is down
let _ = futures::executor::block_on(self.sender.send(()));
}
}
impl SearchQueue {
pub fn new(capacity: usize, paralellism: NonZeroUsize) -> Self {
// Search requests are going to wait until we're available anyway,
// so let's not allocate any RAM and keep a capacity of 1.
let (sender, receiver) = mpsc::channel(1);
tokio::task::spawn(Self::run(capacity, paralellism, receiver));
Self { sender, capacity }
}
/// This function is the main loop, it's in charge on scheduling which search request should execute first and
/// how many should executes at the same time.
///
/// It **must never** panic or exit.
async fn run(
capacity: usize,
parallelism: NonZeroUsize,
mut receive_new_searches: mpsc::Receiver<oneshot::Sender<Permit>>,
) {
let mut queue: Vec<oneshot::Sender<Permit>> = Default::default();
let mut rng: StdRng = StdRng::from_entropy();
let mut searches_running: usize = 0;
// By having a capacity of parallelism we ensures that every time a search finish it can release its RAM asap
let (sender, mut search_finished) = mpsc::channel(parallelism.into());
loop {
tokio::select! {
// biased select because we wants to free up space before trying to register new tasks
biased;
_ = search_finished.recv() => {
searches_running = searches_running.saturating_sub(1);
if !queue.is_empty() {
// Can't panic: the queue wasn't empty thus the range isn't empty.
let remove = rng.gen_range(0..queue.len());
let channel = queue.swap_remove(remove);
let _ = channel.send(Permit { sender: sender.clone() });
}
},
search_request = receive_new_searches.recv() => {
// this unwrap is safe because we're sure the `SearchQueue` still lives somewhere in actix-web
let search_request = search_request.unwrap();
if searches_running < usize::from(parallelism) && queue.is_empty() {
searches_running += 1;
// if the search requests die it's not a hard error on our side
let _ = search_request.send(Permit { sender: sender.clone() });
continue;
} else if capacity == 0 {
// in the very specific case where we have a capacity of zero
// we must refuse the request straight away without going through
// the queue stuff.
drop(search_request);
continue;
} else if queue.len() >= capacity {
let remove = rng.gen_range(0..queue.len());
let thing = queue.swap_remove(remove); // this will drop the channel and notify the search that it won't be processed
drop(thing);
}
queue.push(search_request);
},
}
}
}
/// Returns a search `Permit`.
/// It should be dropped as soon as you've freed all the RAM associated with the search request being processed.
pub async fn try_get_search_permit(&self) -> Result<Permit, MeilisearchHttpError> {
let (sender, receiver) = oneshot::channel();
self.sender.send(sender).await.map_err(|_| MeilisearchHttpError::SearchLimiterIsDown)?;
receiver.await.map_err(|_| MeilisearchHttpError::TooManySearchRequests(self.capacity))
}
/// Returns `Ok(())` if everything seems normal.
/// Returns `Err(MeilisearchHttpError::SearchLimiterIsDown)` if the search limiter seems down.
pub fn health(&self) -> Result<(), MeilisearchHttpError> {
if self.sender.is_closed() {
Err(MeilisearchHttpError::SearchLimiterIsDown)
} else {
Ok(())
}
}
}

5
meilisearch/tests/common/index.rs
View File

@ -328,6 +328,11 @@ impl Index<'_> {
self.service.patch_encoded(url, settings, self.encoder).await
}
pub async fn update_settings_search_cutoff_ms(&self, settings: Value) -> (Value, StatusCode) {
let url = format!("/indexes/{}/settings/search-cutoff-ms", urlencode(self.uid.as_ref()));
self.service.put_encoded(url, settings, self.encoder).await
}
pub async fn delete_settings(&self) -> (Value, StatusCode) {
let url = format!("/indexes/{}/settings", urlencode(self.uid.as_ref()));
self.service.delete(url).await

1
meilisearch/tests/common/mod.rs
View File

@ -16,6 +16,7 @@ pub use server::{default_settings, Server};
pub struct Value(pub serde_json::Value);
impl Value {
#[track_caller]
pub fn uid(&self) -> u64 {
if let Some(uid) = self["uid"].as_u64() {
uid

239
meilisearch/tests/documents/add_documents.rs
View File

@ -1237,8 +1237,8 @@ async fn error_add_documents_missing_document_id() {
}
#[actix_rt::test]
#[ignore] // // TODO: Fix in an other PR: this does not provoke any error.
async fn error_document_field_limit_reached() {
#[should_panic]
async fn error_document_field_limit_reached_in_one_document() {
let server = Server::new().await;
let index = server.index("test");
@ -1246,22 +1246,241 @@ async fn error_document_field_limit_reached() {
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "wow");
for i in 0..65535 {
for i in 0..(u16::MAX as usize + 1) {
let key = i.to_string();
big_object.insert(key, "I am a text!");
}
let documents = json!([big_object]);
let (_response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202");
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"500 Internal Server Error");
index.wait_task(0).await;
let (response, code) = index.get_task(0).await;
snapshot!(code, @"200");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
// Documents without a primary key are not accepted.
snapshot!(json_string!(response, { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" }),
@"");
snapshot!(response,
@r###"
{
"uid": 1,
"indexUid": "test",
"status": "succeeded",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 1
},
"error": null,
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
}
#[actix_rt::test]
async fn error_document_field_limit_reached_over_multiple_documents() {
let server = Server::new().await;
let index = server.index("test");
index.create(Some("id")).await;
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "wow");
for i in 0..(u16::MAX / 2) {
let key = i.to_string();
big_object.insert(key, "I am a text!");
}
let documents = json!([big_object]);
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202 Accepted");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
snapshot!(response,
@r###"
{
"uid": 1,
"indexUid": "test",
"status": "succeeded",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 1
},
"error": null,
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "waw");
for i in (u16::MAX as usize / 2)..(u16::MAX as usize + 1) {
let key = i.to_string();
big_object.insert(key, "I am a text!");
}
let documents = json!([big_object]);
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202 Accepted");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
snapshot!(response,
@r###"
{
"uid": 2,
"indexUid": "test",
"status": "failed",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 0
},
"error": {
"message": "A document cannot contain more than 65,535 fields.",
"code": "max_fields_limit_exceeded",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#max_fields_limit_exceeded"
},
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
}
#[actix_rt::test]
async fn error_document_field_limit_reached_in_one_nested_document() {
let server = Server::new().await;
let index = server.index("test");
index.create(Some("id")).await;
let mut nested = std::collections::HashMap::new();
for i in 0..(u16::MAX as usize + 1) {
let key = i.to_string();
nested.insert(key, "I am a text!");
}
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "wow");
let documents = json!([big_object]);
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202 Accepted");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
// Documents without a primary key are not accepted.
snapshot!(response,
@r###"
{
"uid": 1,
"indexUid": "test",
"status": "succeeded",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 1
},
"error": null,
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
}
#[actix_rt::test]
async fn error_document_field_limit_reached_over_multiple_documents_with_nested_fields() {
let server = Server::new().await;
let index = server.index("test");
index.create(Some("id")).await;
let mut nested = std::collections::HashMap::new();
for i in 0..(u16::MAX / 2) {
let key = i.to_string();
nested.insert(key, "I am a text!");
}
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "wow");
let documents = json!([big_object]);
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202 Accepted");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
snapshot!(response,
@r###"
{
"uid": 1,
"indexUid": "test",
"status": "succeeded",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 1
},
"error": null,
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
let mut nested = std::collections::HashMap::new();
for i in 0..(u16::MAX / 2) {
let key = i.to_string();
nested.insert(key, "I am a text!");
}
let mut big_object = std::collections::HashMap::new();
big_object.insert("id".to_owned(), "wow");
let documents = json!([big_object]);
let (response, code) = index.update_documents(documents, Some("id")).await;
snapshot!(code, @"202 Accepted");
let response = index.wait_task(response.uid()).await;
snapshot!(code, @"202 Accepted");
snapshot!(response,
@r###"
{
"uid": 2,
"indexUid": "test",
"status": "succeeded",
"type": "documentAdditionOrUpdate",
"canceledBy": null,
"details": {
"receivedDocuments": 1,
"indexedDocuments": 1
},
"error": null,
"duration": "[duration]",
"enqueuedAt": "[date]",
"startedAt": "[date]",
"finishedAt": "[date]"
}
"###);
}
#[actix_rt::test]

39
meilisearch/tests/dumps/mod.rs
View File

@ -77,7 +77,8 @@ async fn import_dump_v1_movie_raw() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -238,7 +239,8 @@ async fn import_dump_v1_movie_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -385,7 +387,8 @@ async fn import_dump_v1_rubygems_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -518,7 +521,8 @@ async fn import_dump_v2_movie_raw() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -663,7 +667,8 @@ async fn import_dump_v2_movie_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -807,7 +812,8 @@ async fn import_dump_v2_rubygems_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -940,7 +946,8 @@ async fn import_dump_v3_movie_raw() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1085,7 +1092,8 @@ async fn import_dump_v3_movie_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1229,7 +1237,8 @@ async fn import_dump_v3_rubygems_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1362,7 +1371,8 @@ async fn import_dump_v4_movie_raw() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1507,7 +1517,8 @@ async fn import_dump_v4_movie_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1651,7 +1662,8 @@ async fn import_dump_v4_rubygems_with_settings() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###
);
@ -1895,7 +1907,8 @@ async fn import_dump_v6_containing_experimental_features() {
},
"pagination": {
"maxTotalHits": 1000
}
},
"searchCutoffMs": null
}
"###);

43
meilisearch/tests/search/facet_search.rs
View File

@ -123,6 +123,28 @@ async fn simple_facet_search_with_max_values() {
assert_eq!(dbg!(response)["facetHits"].as_array().unwrap().len(), 1);
}
#[actix_rt::test]
async fn simple_facet_search_by_count_with_max_values() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index
.update_settings_faceting(
json!({ "maxValuesPerFacet": 1, "sortFacetValuesBy": { "*": "count" } }),
)
.await;
index.update_settings_filterable_attributes(json!(["genres"])).await;
index.add_documents(documents, None).await;
index.wait_task(2).await;
let (response, code) =
index.facet_search(json!({"facetName": "genres", "facetQuery": "a"})).await;
assert_eq!(code, 200, "{}", response);
assert_eq!(dbg!(response)["facetHits"].as_array().unwrap().len(), 1);
}
#[actix_rt::test]
async fn non_filterable_facet_search_error() {
let server = Server::new().await;
@ -157,3 +179,24 @@ async fn facet_search_dont_support_words() {
assert_eq!(code, 200, "{}", response);
assert_eq!(response["facetHits"].as_array().unwrap().len(), 0);
}
#[actix_rt::test]
async fn simple_facet_search_with_sort_by_count() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index.update_settings_faceting(json!({ "sortFacetValuesBy": { "*": "count" } })).await;
index.update_settings_filterable_attributes(json!(["genres"])).await;
index.add_documents(documents, None).await;
index.wait_task(2).await;
let (response, code) =
index.facet_search(json!({"facetName": "genres", "facetQuery": "a"})).await;
assert_eq!(code, 200, "{}", response);
let hits = response["facetHits"].as_array().unwrap();
assert_eq!(hits.len(), 2);
assert_eq!(hits[0], json!({ "value": "Action", "count": 3 }));
assert_eq!(hits[1], json!({ "value": "Adventure", "count": 2 }));
}

168
meilisearch/tests/search/hybrid.rs
View File

@ -77,14 +77,57 @@ async fn simple_search() {
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]}},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]}}]"###);
snapshot!(response["semanticHitCount"], @"0");
let (response, code) = index
.search_post(
json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.8}}),
json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.5}, "showRankingScore": true}),
)
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_semanticScore":0.99029034},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_semanticScore":0.97434163},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_semanticScore":0.9472136}]"###);
snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.996969696969697},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.996969696969697},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
snapshot!(response["semanticHitCount"], @"1");
let (response, code) = index
.search_post(
json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.8}, "showRankingScore": true}),
)
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
snapshot!(response["semanticHitCount"], @"3");
}
#[actix_rt::test]
async fn distribution_shift() {
let server = Server::new().await;
let index = index_with_documents(&server, &SIMPLE_SEARCH_DOCUMENTS).await;
let search = json!({"q": "Captain", "vector": [1.0, 1.0], "showRankingScore": true, "hybrid": {"semanticRatio": 1.0}});
let (response, code) = index.search_post(search.clone()).await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
let (response, code) = index
.update_settings(json!({
"embedders": {
"default": {
"distribution": {
"mean": 0.998,
"sigma": 0.01
}
}
}
}))
.await;
snapshot!(code, @"202 Accepted");
let response = server.wait_task(response.uid()).await;
snapshot!(response["details"], @r###"{"embedders":{"default":{"distribution":{"mean":0.998,"sigma":0.01}}}}"###);
let (response, code) = index.search_post(search).await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.19161224365234375},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.1920928955078125e-7},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.1920928955078125e-7}]"###);
}
#[actix_rt::test]
@ -104,10 +147,12 @@ async fn highlighter() {
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}}},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}}}]"###);
snapshot!(response["semanticHitCount"], @"0");
let (response, code) = index
.search_post(json!({"q": "Captain Marvel", "vector": [1.0, 1.0],
"hybrid": {"semanticRatio": 0.8},
"showRankingScore": true,
"attributesToHighlight": [
"desc"
],
@ -116,12 +161,14 @@ async fn highlighter() {
}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_semanticScore":0.99029034},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_semanticScore":0.97434163},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}},"_semanticScore":0.9472136}]"###);
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}},"_rankingScore":0.9472135901451112}]"###);
snapshot!(response["semanticHitCount"], @"3");
// no highlighting on full semantic
let (response, code) = index
.search_post(json!({"q": "Captain Marvel", "vector": [1.0, 1.0],
"hybrid": {"semanticRatio": 1.0},
"showRankingScore": true,
"attributesToHighlight": [
"desc"
],
@ -130,7 +177,8 @@ async fn highlighter() {
}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_semanticScore":0.99029034},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_semanticScore":0.97434163},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}}}]"###);
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}},"_rankingScore":0.9472135901451112}]"###);
snapshot!(response["semanticHitCount"], @"3");
}
#[actix_rt::test]
@ -217,5 +265,115 @@ async fn single_document() {
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"][0], @r###"{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0,"_semanticScore":1.0}"###);
snapshot!(response["hits"][0], @r###"{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0}"###);
snapshot!(response["semanticHitCount"], @"1");
}
#[actix_rt::test]
async fn query_combination() {
let server = Server::new().await;
let index = index_with_documents(&server, &SIMPLE_SEARCH_DOCUMENTS).await;
// search without query and vector, but with hybrid => still placeholder
let (response, code) = index
.search_post(json!({"hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
snapshot!(response["semanticHitCount"], @"null");
// same with a different semantic ratio
let (response, code) = index
.search_post(json!({"hybrid": {"semanticRatio": 0.76}, "showRankingScore": true}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
snapshot!(response["semanticHitCount"], @"null");
// wrong vector dimensions
let (response, code) = index
.search_post(json!({"vector": [1.0, 0.0, 1.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
.await;
snapshot!(code, @"400 Bad Request");
snapshot!(response, @r###"
{
"message": "Invalid vector dimensions: expected: `2`, found: `3`.",
"code": "invalid_vector_dimensions",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#invalid_vector_dimensions"
}
"###);
// full vector
let (response, code) = index
.search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.7773500680923462},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.7236068248748779},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.6581138968467712}]"###);
snapshot!(response["semanticHitCount"], @"3");
// full keyword, without a query
let (response, code) = index
.search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
snapshot!(response["semanticHitCount"], @"null");
// query + vector, full keyword => keyword
let (response, code) = index
.search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true}))
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.996969696969697},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.996969696969697},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.8848484848484849}]"###);
snapshot!(response["semanticHitCount"], @"null");
// query + vector, no hybrid keyword =>
let (response, code) = index
.search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "showRankingScore": true}))
.await;
snapshot!(code, @"400 Bad Request");
snapshot!(response, @r###"
{
"message": "Invalid request: missing `hybrid` parameter when both `q` and `vector` are present.",
"code": "missing_search_hybrid",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#missing_search_hybrid"
}
"###);
// full vector, without a vector => error
let (response, code) = index
.search_post(
json!({"q": "Captain", "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}),
)
.await;
snapshot!(code, @"400 Bad Request");
snapshot!(response, @r###"
{
"message": "Error while generating embeddings: user error: attempt to embed the following text in a configuration where embeddings must be user provided: \"Captain\"",
"code": "vector_embedding_error",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#vector_embedding_error"
}
"###);
// hybrid without a vector => full keyword
let (response, code) = index
.search_post(
json!({"q": "Planet", "hybrid": {"semanticRatio": 0.99}, "showRankingScore": true}),
)
.await;
snapshot!(code, @"200 OK");
snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.9848484848484848}]"###);
snapshot!(response["semanticHitCount"], @"0");
}

206
meilisearch/tests/search/mod.rs
View File

@ -10,6 +10,7 @@ mod hybrid;
mod multi;
mod pagination;
mod restrict_searchable;
mod search_queue;
use once_cell::sync::Lazy;
@ -184,6 +185,110 @@ async fn phrase_search_with_stop_word() {
.await;
}
#[actix_rt::test]
async fn negative_phrase_search() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index.add_documents(documents, None).await;
index.wait_task(0).await;
index
.search(json!({"q": "-\"train your dragon\"" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 4);
assert_eq!(hits[0]["id"], "287947");
assert_eq!(hits[1]["id"], "299537");
assert_eq!(hits[2]["id"], "522681");
assert_eq!(hits[3]["id"], "450465");
})
.await;
}
#[actix_rt::test]
async fn negative_word_search() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index.add_documents(documents, None).await;
index.wait_task(0).await;
index
.search(json!({"q": "-escape" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 4);
assert_eq!(hits[0]["id"], "287947");
assert_eq!(hits[1]["id"], "299537");
assert_eq!(hits[2]["id"], "166428");
assert_eq!(hits[3]["id"], "450465");
})
.await;
// Everything that contains derivates of escape but not escape: nothing
index
.search(json!({"q": "-escape escape" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 0);
})
.await;
}
#[actix_rt::test]
async fn non_negative_search() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index.add_documents(documents, None).await;
index.wait_task(0).await;
index
.search(json!({"q": "- escape" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 1);
assert_eq!(hits[0]["id"], "522681");
})
.await;
index
.search(json!({"q": "- \"train your dragon\"" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 1);
assert_eq!(hits[0]["id"], "166428");
})
.await;
}
#[actix_rt::test]
async fn negative_special_cases_search() {
let server = Server::new().await;
let index = server.index("test");
let documents = DOCUMENTS.clone();
index.add_documents(documents, None).await;
index.wait_task(0).await;
index.update_settings(json!({"synonyms": { "escape": ["glass"] }})).await;
index.wait_task(1).await;
// There is a synonym for escape -> glass but we don't want "escape", only the derivates: glass
index
.search(json!({"q": "-escape escape" }), |response, code| {
assert_eq!(code, 200, "{}", response);
let hits = response["hits"].as_array().unwrap();
assert_eq!(hits.len(), 1);
assert_eq!(hits[0]["id"], "450465");
})
.await;
}
#[cfg(feature = "default")]
#[actix_rt::test]
async fn test_kanji_language_detection() {
@ -834,6 +939,94 @@ async fn test_score_details() {
.await;
}
#[actix_rt::test]
async fn test_degraded_score_details() {
let server = Server::new().await;
let index = server.index("test");
let documents = NESTED_DOCUMENTS.clone();
index.add_documents(json!(documents), None).await;
// We can't really use anything else than 0ms here; otherwise, the test will get flaky.
let (res, _code) = index.update_settings(json!({ "searchCutoffMs": 0 })).await;
index.wait_task(res.uid()).await;
index
.search(
json!({
"q": "b",
"attributesToRetrieve": ["doggos.name", "cattos"],
"showRankingScoreDetails": true,
}),
|response, code| {
meili_snap::snapshot!(code, @"200 OK");
meili_snap::snapshot!(meili_snap::json_string!(response, { ".processingTimeMs" => "[duration]" }), @r###"
{
"hits": [
{
"doggos": [
{
"name": "bobby"
},
{
"name": "buddy"
}
],
"cattos": "pésti",
"_rankingScoreDetails": {
"skipped": {
"order": 0
}
}
},
{
"doggos": [
{
"name": "gros bill"
}
],
"cattos": [
"simba",
"pestiféré"
],
"_rankingScoreDetails": {
"skipped": {
"order": 0
}
}
},
{
"doggos": [
{
"name": "turbo"
},
{
"name": "fast"
}
],
"cattos": [
"moumoute",
"gomez"
],
"_rankingScoreDetails": {
"skipped": {
"order": 0
}
}
}
],
"query": "b",
"processingTimeMs": "[duration]",
"limit": 20,
"offset": 0,
"estimatedTotalHits": 3
}
"###);
},
)
.await;
}
#[actix_rt::test]
async fn experimental_feature_vector_store() {
let server = Server::new().await;
@ -847,6 +1040,7 @@ async fn experimental_feature_vector_store() {
let (response, code) = index
.search_post(json!({
"vector": [1.0, 2.0, 3.0],
"showRankingScore": true
}))
.await;
meili_snap::snapshot!(code, @"400 Bad Request");
@ -889,6 +1083,7 @@ async fn experimental_feature_vector_store() {
let (response, code) = index
.search_post(json!({
"vector": [1.0, 2.0, 3.0],
"showRankingScore": true,
}))
.await;
@ -906,7 +1101,7 @@ async fn experimental_feature_vector_store() {
3
]
},
"_semanticScore": 1.0
"_rankingScore": 1.0
},
{
"title": "Captain Marvel",
@ -918,7 +1113,7 @@ async fn experimental_feature_vector_store() {
54
]
},
"_semanticScore": 0.9129112
"_rankingScore": 0.9129111766815186
},
{
"title": "Gläss",
@ -930,7 +1125,7 @@ async fn experimental_feature_vector_store() {
90
]
},
"_semanticScore": 0.8106413
"_rankingScore": 0.8106412887573242
},
{
"title": "How to Train Your Dragon: The Hidden World",
@ -942,7 +1137,7 @@ async fn experimental_feature_vector_store() {
32
]
},
"_semanticScore": 0.74120104
"_rankingScore": 0.7412010431289673
},
{
"title": "Escape Room",
@ -953,7 +1148,8 @@ async fn experimental_feature_vector_store() {
-23,
32
]
}
},
"_rankingScore": 0.6972063183784485
}
]
"###);

184
meilisearch/tests/search/search_queue.rs Normal file
View File

@ -0,0 +1,184 @@
use std::num::NonZeroUsize;
use std::sync::Arc;
use std::time::Duration;
use actix_web::ResponseError;
use meili_snap::snapshot;
use meilisearch::search_queue::SearchQueue;
#[actix_rt::test]
async fn search_queue_register() {
let queue = SearchQueue::new(4, NonZeroUsize::new(2).unwrap());
// First, use all the cores
let permit1 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
let _permit2 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
// If we free one spot we should be able to register one new search
drop(permit1);
let permit3 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
// And again
drop(permit3);
let _permit4 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
}
#[actix_rt::test]
async fn wait_till_cores_are_available() {
let queue = Arc::new(SearchQueue::new(4, NonZeroUsize::new(1).unwrap()));
// First, use all the cores
let permit1 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
let ret = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit()).await;
assert!(ret.is_err(), "The capacity is full, we should not get a permit");
let q = queue.clone();
let task = tokio::task::spawn(async move { q.try_get_search_permit().await });
// after dropping a permit the previous task should be able to finish
drop(permit1);
let _permit2 = tokio::time::timeout(Duration::from_secs(1), task)
.await
.expect("I should get a permit straight away")
.unwrap();
}
#[actix_rt::test]
async fn refuse_search_requests_when_queue_is_full() {
let queue = Arc::new(SearchQueue::new(1, NonZeroUsize::new(1).unwrap()));
// First, use the whole capacity of the
let _permit1 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
let q = queue.clone();
let permit2 = tokio::task::spawn(async move { q.try_get_search_permit().await });
// Here the queue is full. By registering two new search requests the permit 2 and 3 should be thrown out
let q = queue.clone();
let _permit3 = tokio::task::spawn(async move { q.try_get_search_permit().await });
let permit2 = tokio::time::timeout(Duration::from_secs(1), permit2)
.await
.expect("I should get a result straight away")
.unwrap(); // task should end successfully
let err = meilisearch_types::error::ResponseError::from(permit2.unwrap_err());
let http_response = err.error_response();
let mut headers: Vec<_> = http_response
.headers()
.iter()
.map(|(name, value)| (name.to_string(), value.to_str().unwrap().to_string()))
.collect();
headers.sort();
snapshot!(format!("{headers:?}"), @r###"[("content-type", "application/json"), ("retry-after", "10")]"###);
let err = serde_json::to_string_pretty(&err).unwrap();
snapshot!(err, @r###"
{
"message": "Too many search requests running at the same time: 1. Retry after 10s.",
"code": "too_many_search_requests",
"type": "system",
"link": "https://docs.meilisearch.com/errors#too_many_search_requests"
}
"###);
}
#[actix_rt::test]
async fn search_request_crashes_while_holding_permits() {
let queue = Arc::new(SearchQueue::new(1, NonZeroUsize::new(1).unwrap()));
let (send, recv) = tokio::sync::oneshot::channel();
// This first request take a cpu
let q = queue.clone();
tokio::task::spawn(async move {
let _permit = q.try_get_search_permit().await.unwrap();
recv.await.unwrap();
panic!("oops an unexpected crash happened")
});
// This second request waits in the queue till the first request finishes
let q = queue.clone();
let task = tokio::task::spawn(async move {
let _permit = q.try_get_search_permit().await.unwrap();
});
// By sending something in the channel the request holding a CPU will panic and should lose its permit
send.send(()).unwrap();
// Then the second request should be able to process and finishes correctly without panic
tokio::time::timeout(Duration::from_secs(1), task)
.await
.expect("I should get a permit straight away")
.unwrap();
// I should even be able to take second permit here
let _permit1 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
}
#[actix_rt::test]
async fn works_with_capacity_of_zero() {
let queue = Arc::new(SearchQueue::new(0, NonZeroUsize::new(1).unwrap()));
// First, use the whole capacity of the
let permit1 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
// then we should get an error if we try to register a second search request.
let permit2 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a result straight away");
let err = meilisearch_types::error::ResponseError::from(permit2.unwrap_err());
let http_response = err.error_response();
let mut headers: Vec<_> = http_response
.headers()
.iter()
.map(|(name, value)| (name.to_string(), value.to_str().unwrap().to_string()))
.collect();
headers.sort();
snapshot!(format!("{headers:?}"), @r###"[("content-type", "application/json"), ("retry-after", "10")]"###);
let err = serde_json::to_string_pretty(&err).unwrap();
snapshot!(err, @r###"
{
"message": "Too many search requests running at the same time: 0. Retry after 10s.",
"code": "too_many_search_requests",
"type": "system",
"link": "https://docs.meilisearch.com/errors#too_many_search_requests"
}
"###);
drop(permit1);
// After dropping the first permit we should be able to get a new permit
let _permit3 = tokio::time::timeout(Duration::from_secs(1), queue.try_get_search_permit())
.await
.expect("I should get a permit straight away")
.unwrap();
}

28
meilisearch/tests/settings/errors.rs
View File

@ -337,3 +337,31 @@ async fn settings_bad_pagination() {
}
"###);
}
#[actix_rt::test]
async fn settings_bad_search_cutoff_ms() {
let server = Server::new().await;
let index = server.index("test");
let (response, code) = index.update_settings(json!({ "searchCutoffMs": "doggo" })).await;
snapshot!(code, @"400 Bad Request");
snapshot!(json_string!(response), @r###"
{
"message": "Invalid value type at `.searchCutoffMs`: expected a positive integer, but found a string: `\"doggo\"`",
"code": "invalid_settings_search_cutoff_ms",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#invalid_settings_search_cutoff_ms"
}
"###);
let (response, code) = index.update_settings_search_cutoff_ms(json!("doggo")).await;
snapshot!(code, @"400 Bad Request");
snapshot!(json_string!(response), @r###"
{
"message": "Invalid value type: expected a positive integer, but found a string: `\"doggo\"`",
"code": "invalid_settings_search_cutoff_ms",
"type": "invalid_request",
"link": "https://docs.meilisearch.com/errors#invalid_settings_search_cutoff_ms"
}
"###);
}

141
meilisearch/tests/settings/get_settings.rs
View File

@ -35,6 +35,7 @@ static DEFAULT_SETTINGS_VALUES: Lazy<HashMap<&'static str, Value>> = Lazy::new(|
"maxTotalHits": json!(1000),
}),
);
map.insert("search_cutoff_ms", json!(null));
map
});
@ -49,12 +50,12 @@ async fn get_settings_unexisting_index() {
async fn get_settings() {
let server = Server::new().await;
let index = server.index("test");
index.create(None).await;
index.wait_task(0).await;
let (response, _code) = index.create(None).await;
index.wait_task(response.uid()).await;
let (response, code) = index.settings().await;
assert_eq!(code, 200);
let settings = response.as_object().unwrap();
assert_eq!(settings.keys().len(), 15);
assert_eq!(settings.keys().len(), 16);
assert_eq!(settings["displayedAttributes"], json!(["*"]));
assert_eq!(settings["searchableAttributes"], json!(["*"]));
assert_eq!(settings["filterableAttributes"], json!([]));
@ -84,6 +85,137 @@ async fn get_settings() {
})
);
assert_eq!(settings["proximityPrecision"], json!("byWord"));
assert_eq!(settings["searchCutoffMs"], json!(null));
}
#[actix_rt::test]
async fn secrets_are_hidden_in_settings() {
let server = Server::new().await;
let (response, code) = server.set_features(json!({"vectorStore": true})).await;
meili_snap::snapshot!(code, @"200 OK");
meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
{
"vectorStore": true,
"metrics": false,
"logsRoute": false,
"exportPuffinReports": false
}
"###);
let index = server.index("test");
let (response, _code) = index.create(None).await;
index.wait_task(response.uid()).await;
let (response, code) = index
.update_settings(json!({
"embedders": {
"default": {
"source": "rest",
"url": "https://localhost:7777",
"apiKey": "My super secret value you will never guess"
}
}
}))
.await;
meili_snap::snapshot!(code, @"202 Accepted");
meili_snap::snapshot!(meili_snap::json_string!(response, { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" }),
@r###"
{
"taskUid": 1,
"indexUid": "test",
"status": "enqueued",
"type": "settingsUpdate",
"enqueuedAt": "[date]"
}
"###);
let settings_update_uid = response.uid();
index.wait_task(settings_update_uid).await;
let (response, code) = index.settings().await;
meili_snap::snapshot!(code, @"200 OK");
meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
{
"displayedAttributes": [
"*"
],
"searchableAttributes": [
"*"
],
"filterableAttributes": [],
"sortableAttributes": [],
"rankingRules": [
"words",
"typo",
"proximity",
"attribute",
"sort",
"exactness"
],
"stopWords": [],
"nonSeparatorTokens": [],
"separatorTokens": [],
"dictionary": [],
"synonyms": {},
"distinctAttribute": null,
"proximityPrecision": "byWord",
"typoTolerance": {
"enabled": true,
"minWordSizeForTypos": {
"oneTypo": 5,
"twoTypos": 9
},
"disableOnWords": [],
"disableOnAttributes": []
},
"faceting": {
"maxValuesPerFacet": 100,
"sortFacetValuesBy": {
"*": "alpha"
}
},
"pagination": {
"maxTotalHits": 1000
},
"embedders": {
"default": {
"source": "rest",
"apiKey": "My suXXXXXX...",
"documentTemplate": "{% for field in fields %} {{ field.name }}: {{ field.value }}\n{% endfor %}",
"url": "https://localhost:7777",
"query": null,
"inputField": [
"input"
],
"pathToEmbeddings": [
"data"
],
"embeddingObject": [
"embedding"
],
"inputType": "text"
}
},
"searchCutoffMs": null
}
"###);
let (response, code) = server.get_task(settings_update_uid).await;
meili_snap::snapshot!(code, @"200 OK");
meili_snap::snapshot!(meili_snap::json_string!(response["details"]), @r###"
{
"embedders": {
"default": {
"source": "rest",
"apiKey": "My suXXXXXX...",
"url": "https://localhost:7777"
}
}
}
"###);
}
#[actix_rt::test]
@ -285,7 +417,8 @@ test_setting_routes!(
ranking_rules put,
synonyms put,
pagination patch,
faceting patch
faceting patch,
search_cutoff_ms put
);
#[actix_rt::test]

6
meilitool/src/main.rs
View File

@ -291,7 +291,11 @@ fn export_a_dump(
}
// 4.2. Dump the settings
let settings = meilisearch_types::settings::settings(&index, &rtxn)?;
let settings = meilisearch_types::settings::settings(
&index,
&rtxn,
meilisearch_types::settings::SecretPolicy::RevealSecrets,
)?;
index_dumper.settings(&settings)?;
count += 1;
}

18
milli/Cargo.toml
View File

@ -17,7 +17,7 @@ bincode = "1.3.3"
bstr = "1.9.0"
bytemuck = { version = "1.14.0", features = ["extern_crate_alloc"] }
byteorder = "1.5.0"
charabia = { version = "0.8.7", default-features = false }
charabia = { version = "0.8.8", default-features = false }
concat-arrays = "0.1.2"
crossbeam-channel = "0.5.11"
deserr = "0.6.1"
@ -71,26 +71,22 @@ itertools = "0.11.0"
puffin = "0.16.0"
csv = "1.3.0"
candle-core = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
candle-transformers = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
candle-nn = { git = "https://github.com/huggingface/candle.git", version = "0.3.1" }
tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.14.1", version = "0.14.1", default_features = false, features = [
candle-core = { version = "0.4.1" }
candle-transformers = { version = "0.4.1" }
candle-nn = { version = "0.4.1" }
tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.15.2", version = "0.15.2", default_features = false, features = [
"onig",
] }
hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls", default_features = false, features = [
"online",
] }
tokio = { version = "1.35.1", features = ["rt"] }
futures = "0.3.30"
reqwest = { version = "0.11.23", features = [
"rustls-tls",
"json",
], default-features = false }
tiktoken-rs = "0.5.8"
liquid = "0.26.4"
arroy = "0.2.0"
rand = "0.8.5"
tracing = "0.1.40"
ureq = { version = "2.9.6", features = ["json"] }
url = "2.5.0"
[dev-dependencies]
mimalloc = { version = "0.1.39", default-features = false }

3
milli/examples/search.rs
View File

@ -6,7 +6,7 @@ use std::time::Instant;
use heed::EnvOpenOptions;
use milli::{
execute_search, filtered_universe, DefaultSearchLogger, GeoSortStrategy, Index, SearchContext,
SearchLogger, TermsMatchingStrategy,
SearchLogger, TermsMatchingStrategy, TimeBudget,
};
#[global_allocator]
@ -65,6 +65,7 @@ fn main() -> Result<(), Box<dyn Error>> {
None,
&mut DefaultSearchLogger,
logger,
TimeBudget::max(),
)?;
if let Some((logger, dir)) = detailed_logger {
logger.finish(&mut ctx, Path::new(dir))?;

4
milli/src/error.rs
View File

@ -196,7 +196,7 @@ only composed of alphanumeric characters (a-z A-Z 0-9), hyphens (-) and undersco
InvalidPromptForEmbeddings(String, crate::prompt::error::NewPromptError),
#[error("Too many embedders in the configuration. Found {0}, but limited to 256.")]
TooManyEmbedders(usize),
#[error("Cannot find embedder with name {0}.")]
#[error("Cannot find embedder with name `{0}`.")]
InvalidEmbedder(String),
#[error("Too many vectors for document with id {0}: found {1}, but limited to 256.")]
TooManyVectors(String, usize),
@ -243,6 +243,8 @@ only composed of alphanumeric characters (a-z A-Z 0-9), hyphens (-) and undersco
},
#[error("`.embedders.{embedder_name}.dimensions`: `dimensions` cannot be zero")]
InvalidSettingsDimensions { embedder_name: String },
#[error("`.embedders.{embedder_name}.url`: could not parse `{url}`: {inner_error}")]
InvalidUrl { embedder_name: String, inner_error: url::ParseError, url: String },
}
impl From<crate::vector::Error> for Error {

37
milli/src/index.rs
View File

@ -20,13 +20,13 @@ use crate::heed_codec::facet::{
use crate::heed_codec::{
BEU16StrCodec, FstSetCodec, ScriptLanguageCodec, StrBEU16Codec, StrRefCodec,
};
use crate::order_by_map::OrderByMap;
use crate::proximity::ProximityPrecision;
use crate::vector::EmbeddingConfig;
use crate::{
default_criteria, CboRoaringBitmapCodec, Criterion, DocumentId, ExternalDocumentsIds,
FacetDistribution, FieldDistribution, FieldId, FieldIdWordCountCodec, GeoPoint, ObkvCodec,
OrderBy, Result, RoaringBitmapCodec, RoaringBitmapLenCodec, Search, U8StrStrCodec, BEU16,
BEU32, BEU64,
Result, RoaringBitmapCodec, RoaringBitmapLenCodec, Search, U8StrStrCodec, BEU16, BEU32, BEU64,
};
pub const DEFAULT_MIN_WORD_LEN_ONE_TYPO: u8 = 5;
@ -67,6 +67,7 @@ pub mod main_key {
pub const PAGINATION_MAX_TOTAL_HITS: &str = "pagination-max-total-hits";
pub const PROXIMITY_PRECISION: &str = "proximity-precision";
pub const EMBEDDING_CONFIGS: &str = "embedding_configs";
pub const SEARCH_CUTOFF: &str = "search_cutoff";
}
pub mod db_name {
@ -1115,7 +1116,7 @@ impl Index {
/* words prefixes fst */
/// Writes the FST which is the words prefixes dictionnary of the engine.
/// Writes the FST which is the words prefixes dictionary of the engine.
pub(crate) fn put_words_prefixes_fst<A: AsRef<[u8]>>(
&self,
wtxn: &mut RwTxn,
@ -1128,7 +1129,7 @@ impl Index {
)
}
/// Returns the FST which is the words prefixes dictionnary of the engine.
/// Returns the FST which is the words prefixes dictionary of the engine.
pub fn words_prefixes_fst<'t>(&self, rtxn: &'t RoTxn) -> Result<fst::Set<Cow<'t, [u8]>>> {
match self.main.remap_types::<Str, Bytes>().get(rtxn, main_key::WORDS_PREFIXES_FST_KEY)? {
Some(bytes) => Ok(fst::Set::new(bytes)?.map_data(Cow::Borrowed)?),
@ -1373,21 +1374,19 @@ impl Index {
self.main.remap_key_type::<Str>().delete(txn, main_key::MAX_VALUES_PER_FACET)
}
pub fn sort_facet_values_by(&self, txn: &RoTxn) -> heed::Result<HashMap<String, OrderBy>> {
let mut orders = self
pub fn sort_facet_values_by(&self, txn: &RoTxn) -> heed::Result<OrderByMap> {
let orders = self
.main
.remap_types::<Str, SerdeJson<HashMap<String, OrderBy>>>()
.remap_types::<Str, SerdeJson<OrderByMap>>()
.get(txn, main_key::SORT_FACET_VALUES_BY)?
.unwrap_or_default();
// Insert the default ordering if it is not already overwritten by the user.
orders.entry("*".to_string()).or_insert(OrderBy::Lexicographic);
Ok(orders)
}
pub(crate) fn put_sort_facet_values_by(
&self,
txn: &mut RwTxn,
val: &HashMap<String, OrderBy>,
val: &OrderByMap,
) -> heed::Result<()> {
self.main.remap_types::<Str, SerdeJson<_>>().put(txn, main_key::SORT_FACET_VALUES_BY, &val)
}
@ -1500,12 +1499,16 @@ impl Index {
.unwrap_or_default())
}
pub fn default_embedding_name(&self, rtxn: &RoTxn<'_>) -> Result<String> {
let configs = self.embedding_configs(rtxn)?;
Ok(match configs.as_slice() {
[(ref first_name, _)] => first_name.clone(),
_ => "default".to_owned(),
})
pub(crate) fn put_search_cutoff(&self, wtxn: &mut RwTxn<'_>, cutoff: u64) -> heed::Result<()> {
self.main.remap_types::<Str, BEU64>().put(wtxn, main_key::SEARCH_CUTOFF, &cutoff)
}
pub fn search_cutoff(&self, rtxn: &RoTxn<'_>) -> Result<Option<u64>> {
Ok(self.main.remap_types::<Str, BEU64>().get(rtxn, main_key::SEARCH_CUTOFF)?)
}
pub(crate) fn delete_search_cutoff(&self, wtxn: &mut RwTxn<'_>) -> heed::Result<bool> {
self.main.remap_key_type::<Str>().delete(wtxn, main_key::SEARCH_CUTOFF)
}
}
@ -2423,6 +2426,8 @@ pub(crate) mod tests {
candidates: _,
document_scores: _,
mut documents_ids,
degraded: _,
used_negative_operator: _,
} = search.execute().unwrap();
let primary_key_id = index.fields_ids_map(&rtxn).unwrap().id("primary_key").unwrap();
documents_ids.sort_unstable();

75
milli/src/lib.rs
View File

@ -16,6 +16,7 @@ pub mod facet;
mod fields_ids_map;
pub mod heed_codec;
pub mod index;
pub mod order_by_map;
pub mod prompt;
pub mod proximity;
pub mod score_details;
@ -29,6 +30,7 @@ pub mod snapshot_tests;
use std::collections::{BTreeMap, HashMap};
use std::convert::{TryFrom, TryInto};
use std::fmt;
use std::hash::BuildHasherDefault;
use charabia::normalizer::{CharNormalizer, CompatibilityDecompositionNormalizer};
@ -56,10 +58,10 @@ pub use self::heed_codec::{
UncheckedU8StrStrCodec,
};
pub use self::index::Index;
pub use self::search::facet::{FacetValueHit, SearchForFacetValues};
pub use self::search::{
FacetDistribution, FacetValueHit, Filter, FormatOptions, MatchBounds, MatcherBuilder,
MatchingWords, OrderBy, Search, SearchForFacetValues, SearchResult, TermsMatchingStrategy,
DEFAULT_VALUES_PER_FACET,
FacetDistribution, Filter, FormatOptions, MatchBounds, MatcherBuilder, MatchingWords, OrderBy,
Search, SearchResult, SemanticSearch, TermsMatchingStrategy, DEFAULT_VALUES_PER_FACET,
};
pub type Result<T> = std::result::Result<T, error::Error>;
@ -103,6 +105,73 @@ pub const MAX_WORD_LENGTH: usize = MAX_LMDB_KEY_LENGTH / 2;
pub const MAX_POSITION_PER_ATTRIBUTE: u32 = u16::MAX as u32 + 1;
#[derive(Clone)]
pub struct TimeBudget {
started_at: std::time::Instant,
budget: std::time::Duration,
/// When testing the time budget, ensuring we did more than iteration of the bucket sort can be useful.
/// But to avoid being flaky, the only option is to add the ability to stop after a specific number of calls instead of a `Duration`.
#[cfg(test)]
stop_after: Option<(std::sync::Arc<std::sync::atomic::AtomicUsize>, usize)>,
}
impl fmt::Debug for TimeBudget {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
f.debug_struct("TimeBudget")
.field("started_at", &self.started_at)
.field("budget", &self.budget)
.field("left", &(self.budget - self.started_at.elapsed()))
.finish()
}
}
impl Default for TimeBudget {
fn default() -> Self {
Self::new(std::time::Duration::from_millis(150))
}
}
impl TimeBudget {
pub fn new(budget: std::time::Duration) -> Self {
Self {
started_at: std::time::Instant::now(),
budget,
#[cfg(test)]
stop_after: None,
}
}
pub fn max() -> Self {
Self::new(std::time::Duration::from_secs(u64::MAX))
}
#[cfg(test)]
pub fn with_stop_after(mut self, stop_after: usize) -> Self {
use std::sync::atomic::AtomicUsize;
use std::sync::Arc;
self.stop_after = Some((Arc::new(AtomicUsize::new(0)), stop_after));
self
}
pub fn exceeded(&self) -> bool {
#[cfg(test)]
if let Some((current, stop_after)) = &self.stop_after {
let current = current.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
if current >= *stop_after {
return true;
} else {
// if a number has been specified then we ignore entirely the time budget
return false;
}
}
self.started_at.elapsed() > self.budget
}
}
// Convert an absolute word position into a relative position.
// Return the field id of the attribute related to the absolute position
// and the relative position in the attribute.

57
milli/src/order_by_map.rs Normal file
View File

@ -0,0 +1,57 @@
use std::collections::{hash_map, HashMap};
use std::iter::FromIterator;
use serde::{Deserialize, Deserializer, Serialize};
use crate::OrderBy;
#[derive(Serialize)]
pub struct OrderByMap(HashMap<String, OrderBy>);
impl OrderByMap {
pub fn get(&self, key: impl AsRef<str>) -> OrderBy {
self.0
.get(key.as_ref())
.copied()
.unwrap_or_else(|| self.0.get("*").copied().unwrap_or_default())
}
pub fn insert(&mut self, key: String, value: OrderBy) -> Option<OrderBy> {
self.0.insert(key, value)
}
}
impl Default for OrderByMap {
fn default() -> Self {
let mut map = HashMap::new();
map.insert("*".to_string(), OrderBy::Lexicographic);
OrderByMap(map)
}
}
impl FromIterator<(String, OrderBy)> for OrderByMap {
fn from_iter<T: IntoIterator<Item = (String, OrderBy)>>(iter: T) -> Self {
OrderByMap(iter.into_iter().collect())
}
}
impl IntoIterator for OrderByMap {
type Item = (String, OrderBy);
type IntoIter = hash_map::IntoIter<String, OrderBy>;
fn into_iter(self) -> Self::IntoIter {
self.0.into_iter()
}
}
impl<'de> Deserialize<'de> for OrderByMap {
fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
where
D: Deserializer<'de>,
{
let mut map = Deserialize::deserialize(deserializer).map(OrderByMap)?;
// Insert the default ordering if it is not already overwritten by the user.
map.0.entry("*".to_string()).or_insert(OrderBy::default());
Ok(map)
}
}

26
milli/src/score_details.rs
View File

@ -17,6 +17,9 @@ pub enum ScoreDetails {
Sort(Sort),
Vector(Vector),
GeoSort(GeoSort),
/// Returned when we don't have the time to finish applying all the subsequent ranking-rules
Skipped,
}
#[derive(Clone, Copy)]
@ -50,6 +53,7 @@ impl ScoreDetails {
ScoreDetails::Sort(_) => None,
ScoreDetails::GeoSort(_) => None,
ScoreDetails::Vector(_) => None,
ScoreDetails::Skipped => Some(Rank { rank: 0, max_rank: 1 }),
}
}
@ -94,9 +98,10 @@ impl ScoreDetails {
ScoreDetails::ExactWords(e) => RankOrValue::Rank(e.rank()),
ScoreDetails::Sort(sort) => RankOrValue::Sort(sort),
ScoreDetails::GeoSort(geosort) => RankOrValue::GeoSort(geosort),
ScoreDetails::Vector(vector) => RankOrValue::Score(
vector.value_similarity.as_ref().map(|(_, s)| *s as f64).unwrap_or(0.0f64),
),
ScoreDetails::Vector(vector) => {
RankOrValue::Score(vector.similarity.as_ref().map(|s| *s as f64).unwrap_or(0.0f64))
}
ScoreDetails::Skipped => RankOrValue::Rank(Rank { rank: 0, max_rank: 1 }),
}
}
@ -244,16 +249,18 @@ impl ScoreDetails {
order += 1;
}
ScoreDetails::Vector(s) => {
let vector = format!("vectorSort({:?})", s.target_vector);
let value = s.value_similarity.as_ref().map(|(v, _)| v);
let similarity = s.value_similarity.as_ref().map(|(_, s)| s);
let similarity = s.similarity.as_ref();
let details = serde_json::json!({
"order": order,
"value": value,
"similarity": similarity,
});
details_map.insert(vector, details);
details_map.insert("vectorSort".into(), details);
order += 1;
}
ScoreDetails::Skipped => {
details_map
.insert("skipped".to_string(), serde_json::json!({ "order": order }));
order += 1;
}
}
@ -484,8 +491,7 @@ impl PartialOrd for GeoSort {
#[derive(Debug, Clone, PartialEq, PartialOrd)]
pub struct Vector {
pub target_vector: Vec<f32>,
pub value_similarity: Option<(Vec<f32>, f32)>,
pub similarity: Option<f32>,
}
impl GeoSort {

4
milli/src/search/facet/facet_range_search.rs
View File

@ -168,7 +168,7 @@ impl<'t, 'b, 'bitmap> FacetRangeSearch<'t, 'b, 'bitmap> {
}
// should we stop?
// We should if the the search range doesn't include any
// We should if the search range doesn't include any
// element from the previous key or its successors
let should_stop = {
match self.right {
@ -232,7 +232,7 @@ impl<'t, 'b, 'bitmap> FacetRangeSearch<'t, 'b, 'bitmap> {
}
// should we stop?
// We should if the the search range doesn't include any
// We should if the search range doesn't include any
// element from the previous key or its successors
let should_stop = {
match self.right {

3
milli/src/search/facet/mod.rs
View File

@ -6,15 +6,18 @@ use roaring::RoaringBitmap;
pub use self::facet_distribution::{FacetDistribution, OrderBy, DEFAULT_VALUES_PER_FACET};
pub use self::filter::{BadGeoError, Filter};
pub use self::search::{FacetValueHit, SearchForFacetValues};
use crate::heed_codec::facet::{FacetGroupKeyCodec, FacetGroupValueCodec, OrderedF64Codec};
use crate::heed_codec::BytesRefCodec;
use crate::{Index, Result};
mod facet_distribution;
mod facet_distribution_iter;
mod facet_range_search;
mod facet_sort_ascending;
mod facet_sort_descending;
mod filter;
mod search;
fn facet_extreme_value<'t>(
mut extreme_it: impl Iterator<Item = heed::Result<(RoaringBitmap, &'t [u8])>> + 't,

332
milli/src/search/facet/search.rs Normal file
View File

@ -0,0 +1,332 @@
use std::cmp::{Ordering, Reverse};
use std::collections::BinaryHeap;
use std::ops::ControlFlow;
use charabia::normalizer::NormalizerOption;
use charabia::Normalize;
use fst::automaton::{Automaton, Str};
use fst::{IntoStreamer, Streamer};
use roaring::RoaringBitmap;
use tracing::error;
use crate::error::UserError;
use crate::heed_codec::facet::{FacetGroupKey, FacetGroupValue};
use crate::search::build_dfa;
use crate::{DocumentId, FieldId, OrderBy, Result, Search};
/// The maximum number of values per facet returned by the facet search route.
const DEFAULT_MAX_NUMBER_OF_VALUES_PER_FACET: usize = 100;
pub struct SearchForFacetValues<'a> {
query: Option<String>,
facet: String,
search_query: Search<'a>,
max_values: usize,
is_hybrid: bool,
}
impl<'a> SearchForFacetValues<'a> {
pub fn new(
facet: String,
search_query: Search<'a>,
is_hybrid: bool,
) -> SearchForFacetValues<'a> {
SearchForFacetValues {
query: None,
facet,
search_query,
max_values: DEFAULT_MAX_NUMBER_OF_VALUES_PER_FACET,
is_hybrid,
}
}
pub fn query(&mut self, query: impl Into<String>) -> &mut Self {
self.query = Some(query.into());
self
}
pub fn max_values(&mut self, max: usize) -> &mut Self {
self.max_values = max;
self
}
fn one_original_value_of(
&self,
field_id: FieldId,
facet_str: &str,
any_docid: DocumentId,
) -> Result<Option<String>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let key: (FieldId, _, &str) = (field_id, any_docid, facet_str);
Ok(index.field_id_docid_facet_strings.get(rtxn, &key)?.map(|v| v.to_owned()))
}
pub fn execute(&self) -> Result<Vec<FacetValueHit>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let filterable_fields = index.filterable_fields(rtxn)?;
if !filterable_fields.contains(&self.facet) {
let (valid_fields, hidden_fields) =
index.remove_hidden_fields(rtxn, filterable_fields)?;
return Err(UserError::InvalidFacetSearchFacetName {
field: self.facet.clone(),
valid_fields,
hidden_fields,
}
.into());
}
let fields_ids_map = index.fields_ids_map(rtxn)?;
let fid = match fields_ids_map.id(&self.facet) {
Some(fid) => fid,
// we return an empty list of results when the attribute has been
// set as filterable but no document contains this field (yet).
None => return Ok(Vec::new()),
};
let fst = match self.search_query.index.facet_id_string_fst.get(rtxn, &fid)? {
Some(fst) => fst,
None => return Ok(Vec::new()),
};
let search_candidates = self.search_query.execute_for_candidates(
self.is_hybrid
|| self
.search_query
.semantic
.as_ref()
.and_then(|semantic| semantic.vector.as_ref())
.is_some(),
)?;
let mut results = match index.sort_facet_values_by(rtxn)?.get(&self.facet) {
OrderBy::Lexicographic => ValuesCollection::by_lexicographic(self.max_values),
OrderBy::Count => ValuesCollection::by_count(self.max_values),
};
match self.query.as_ref() {
Some(query) => {
let options = NormalizerOption { lossy: true, ..Default::default() };
let query = query.normalize(&options);
let query = query.as_ref();
let authorize_typos = self.search_query.index.authorize_typos(rtxn)?;
let field_authorizes_typos =
!self.search_query.index.exact_attributes_ids(rtxn)?.contains(&fid);
if authorize_typos && field_authorizes_typos {
let exact_words_fst = self.search_query.index.exact_words(rtxn)?;
if exact_words_fst.map_or(false, |fst| fst.contains(query)) {
if fst.contains(query) {
self.fetch_original_facets_using_normalized(
fid,
query,
query,
&search_candidates,
&mut results,
)?;
}
} else {
let one_typo = self.search_query.index.min_word_len_one_typo(rtxn)?;
let two_typos = self.search_query.index.min_word_len_two_typos(rtxn)?;
let is_prefix = true;
let automaton = if query.len() < one_typo as usize {
build_dfa(query, 0, is_prefix)
} else if query.len() < two_typos as usize {
build_dfa(query, 1, is_prefix)
} else {
build_dfa(query, 2, is_prefix)
};
let mut stream = fst.search(automaton).into_stream();
while let Some(facet_value) = stream.next() {
let value = std::str::from_utf8(facet_value)?;
if self
.fetch_original_facets_using_normalized(
fid,
value,
query,
&search_candidates,
&mut results,
)?
.is_break()
{
break;
}
}
}
} else {
let automaton = Str::new(query).starts_with();
let mut stream = fst.search(automaton).into_stream();
while let Some(facet_value) = stream.next() {
let value = std::str::from_utf8(facet_value)?;
if self
.fetch_original_facets_using_normalized(
fid,
value,
query,
&search_candidates,
&mut results,
)?
.is_break()
{
break;
}
}
}
}
None => {
let prefix = FacetGroupKey { field_id: fid, level: 0, left_bound: "" };
for result in index.facet_id_string_docids.prefix_iter(rtxn, &prefix)? {
let (FacetGroupKey { left_bound, .. }, FacetGroupValue { bitmap, .. }) =
result?;
let count = search_candidates.intersection_len(&bitmap);
if count != 0 {
let value = self
.one_original_value_of(fid, left_bound, bitmap.min().unwrap())?
.unwrap_or_else(|| left_bound.to_string());
if results.insert(FacetValueHit { value, count }).is_break() {
break;
}
}
}
}
}
Ok(results.into_sorted_vec())
}
fn fetch_original_facets_using_normalized(
&self,
fid: FieldId,
value: &str,
query: &str,
search_candidates: &RoaringBitmap,
results: &mut ValuesCollection,
) -> Result<ControlFlow<()>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let database = index.facet_id_normalized_string_strings;
let key = (fid, value);
let original_strings = match database.get(rtxn, &key)? {
Some(original_strings) => original_strings,
None => {
error!("the facet value is missing from the facet database: {key:?}");
return Ok(ControlFlow::Continue(()));
}
};
for original in original_strings {
let key = FacetGroupKey { field_id: fid, level: 0, left_bound: original.as_str() };
let docids = match index.facet_id_string_docids.get(rtxn, &key)? {
Some(FacetGroupValue { bitmap, .. }) => bitmap,
None => {
error!("the facet value is missing from the facet database: {key:?}");
return Ok(ControlFlow::Continue(()));
}
};
let count = search_candidates.intersection_len(&docids);
if count != 0 {
let value = self
.one_original_value_of(fid, &original, docids.min().unwrap())?
.unwrap_or_else(|| query.to_string());
if results.insert(FacetValueHit { value, count }).is_break() {
break;
}
}
}
Ok(ControlFlow::Continue(()))
}
}
#[derive(Debug, Clone, serde::Serialize, PartialEq)]
pub struct FacetValueHit {
/// The original facet value
pub value: String,
/// The number of documents associated to this facet
pub count: u64,
}
impl PartialOrd for FacetValueHit {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl Ord for FacetValueHit {
fn cmp(&self, other: &Self) -> Ordering {
self.count.cmp(&other.count).then_with(|| self.value.cmp(&other.value))
}
}
impl Eq for FacetValueHit {}
/// A wrapper type that collects the best facet values by
/// lexicographic or number of associated values.
enum ValuesCollection {
/// Keeps the top values according to the lexicographic order.
Lexicographic { max: usize, content: Vec<FacetValueHit> },
/// Keeps the top values according to the number of values associated to them.
///
/// Note that it is a max heap and we need to move the smallest counts
/// at the top to be able to pop them when we reach the max_values limit.
Count { max: usize, content: BinaryHeap<Reverse<FacetValueHit>> },
}
impl ValuesCollection {
pub fn by_lexicographic(max: usize) -> Self {
ValuesCollection::Lexicographic { max, content: Vec::new() }
}
pub fn by_count(max: usize) -> Self {
ValuesCollection::Count { max, content: BinaryHeap::new() }
}
pub fn insert(&mut self, value: FacetValueHit) -> ControlFlow<()> {
match self {
ValuesCollection::Lexicographic { max, content } => {
if content.len() < *max {
content.push(value);
if content.len() < *max {
return ControlFlow::Continue(());
}
}
ControlFlow::Break(())
}
ValuesCollection::Count { max, content } => {
if content.len() == *max {
// Peeking gives us the worst value in the list as
// this is a max-heap and we reversed it.
let Some(mut peek) = content.peek_mut() else { return ControlFlow::Break(()) };
if peek.0.count <= value.count {
// Replace the current worst value in the heap
// with the new one we received that is better.
*peek = Reverse(value);
}
} else {
content.push(Reverse(value));
}
ControlFlow::Continue(())
}
}
}
/// Returns the list of facet values in descending order of, either,
/// count or lexicographic order of the value depending on the type.
pub fn into_sorted_vec(self) -> Vec<FacetValueHit> {
match self {
ValuesCollection::Lexicographic { content, .. } => content.into_iter().collect(),
ValuesCollection::Count { content, .. } => {
// Convert the heap into a vec of hits by removing the Reverse wrapper.
// Hits are already in the right order as they were reversed and there
// are output in ascending order.
content.into_sorted_vec().into_iter().map(|Reverse(hit)| hit).collect()
}
}
}
}

124
milli/src/search/hybrid.rs
View File

@ -4,12 +4,15 @@ use itertools::Itertools;
use roaring::RoaringBitmap;
use crate::score_details::{ScoreDetails, ScoreValue, ScoringStrategy};
use crate::search::SemanticSearch;
use crate::{MatchingWords, Result, Search, SearchResult};
struct ScoreWithRatioResult {
matching_words: MatchingWords,
candidates: RoaringBitmap,
document_scores: Vec<(u32, ScoreWithRatio)>,
degraded: bool,
used_negative_operator: bool,
}
type ScoreWithRatio = (Vec<ScoreDetails>, f32);
@ -49,8 +52,12 @@ fn compare_scores(
order => return order,
}
}
(Some(ScoreValue::Score(_)), Some(_)) => return Ordering::Greater,
(Some(_), Some(ScoreValue::Score(_))) => return Ordering::Less,
(Some(ScoreValue::Score(x)), Some(_)) => {
return if x == 0. { Ordering::Less } else { Ordering::Greater }
}
(Some(_), Some(ScoreValue::Score(x))) => {
return if x == 0. { Ordering::Greater } else { Ordering::Less }
}
// if we have this, we're bad
(Some(ScoreValue::GeoSort(_)), Some(ScoreValue::Sort(_)))
| (Some(ScoreValue::Sort(_)), Some(ScoreValue::GeoSort(_))) => {
@ -72,51 +79,82 @@ impl ScoreWithRatioResult {
matching_words: results.matching_words,
candidates: results.candidates,
document_scores,
degraded: results.degraded,
used_negative_operator: results.used_negative_operator,
}
}
fn merge(left: Self, right: Self, from: usize, length: usize) -> SearchResult {
let mut documents_ids =
Vec::with_capacity(left.document_scores.len() + right.document_scores.len());
let mut document_scores =
Vec::with_capacity(left.document_scores.len() + right.document_scores.len());
fn merge(
vector_results: Self,
keyword_results: Self,
from: usize,
length: usize,
) -> (SearchResult, u32) {
#[derive(Clone, Copy)]
enum ResultSource {
Semantic,
Keyword,
}
let mut semantic_hit_count = 0;
let mut documents_ids = Vec::with_capacity(
vector_results.document_scores.len() + keyword_results.document_scores.len(),
);
let mut document_scores = Vec::with_capacity(
vector_results.document_scores.len() + keyword_results.document_scores.len(),
);
let mut documents_seen = RoaringBitmap::new();
for (docid, (main_score, _sub_score)) in left
for ((docid, (main_score, _sub_score)), source) in vector_results
.document_scores
.into_iter()
.merge_by(right.document_scores.into_iter(), |(_, left), (_, right)| {
// the first value is the one with the greatest score
compare_scores(left, right).is_ge()
})
.zip(std::iter::repeat(ResultSource::Semantic))
.merge_by(
keyword_results
.document_scores
.into_iter()
.zip(std::iter::repeat(ResultSource::Keyword)),
|((_, left), _), ((_, right), _)| {
// the first value is the one with the greatest score
compare_scores(left, right).is_ge()
},
)
// remove documents we already saw
.filter(|(docid, _)| documents_seen.insert(*docid))
.filter(|((docid, _), _)| documents_seen.insert(*docid))
// start skipping **after** the filter
.skip(from)
// take **after** skipping
.take(length)
{
if let ResultSource::Semantic = source {
semantic_hit_count += 1;
}
documents_ids.push(docid);
// TODO: pass both scores to documents_score in some way?
document_scores.push(main_score);
}
SearchResult {
matching_words: right.matching_words,
candidates: left.candidates | right.candidates,
documents_ids,
document_scores,
}
(
SearchResult {
matching_words: keyword_results.matching_words,
candidates: vector_results.candidates | keyword_results.candidates,
documents_ids,
document_scores,
degraded: vector_results.degraded | keyword_results.degraded,
used_negative_operator: vector_results.used_negative_operator
| keyword_results.used_negative_operator,
},
semantic_hit_count,
)
}
}
impl<'a> Search<'a> {
pub fn execute_hybrid(&self, semantic_ratio: f32) -> Result<SearchResult> {
pub fn execute_hybrid(&self, semantic_ratio: f32) -> Result<(SearchResult, Option<u32>)> {
// TODO: find classier way to achieve that than to reset vector and query params
// create separate keyword and semantic searches
let mut search = Search {
query: self.query.clone(),
vector: self.vector.clone(),
filter: self.filter.clone(),
offset: 0,
limit: self.limit + self.offset,
@ -129,25 +167,43 @@ impl<'a> Search<'a> {
exhaustive_number_hits: self.exhaustive_number_hits,
rtxn: self.rtxn,
index: self.index,
distribution_shift: self.distribution_shift,
embedder_name: self.embedder_name.clone(),
semantic: self.semantic.clone(),
time_budget: self.time_budget.clone(),
};
let vector_query = search.vector.take();
let semantic = search.semantic.take();
let keyword_results = search.execute()?;
// skip semantic search if we don't have a vector query (placeholder search)
let Some(vector_query) = vector_query else {
return Ok(keyword_results);
};
// completely skip semantic search if the results of the keyword search are good enough
if self.results_good_enough(&keyword_results, semantic_ratio) {
return Ok(keyword_results);
return Ok((keyword_results, Some(0)));
}
search.vector = Some(vector_query);
search.query = None;
// no vector search against placeholder search
let Some(query) = search.query.take() else {
return Ok((keyword_results, Some(0)));
};
// no embedder, no semantic search
let Some(SemanticSearch { vector, embedder_name, embedder }) = semantic else {
return Ok((keyword_results, Some(0)));
};
let vector_query = match vector {
Some(vector_query) => vector_query,
None => {
// attempt to embed the vector
match embedder.embed_one(query) {
Ok(embedding) => embedding,
Err(error) => {
tracing::error!(error=%error, "Embedding failed");
return Ok((keyword_results, Some(0)));
}
}
}
};
search.semantic =
Some(SemanticSearch { vector: Some(vector_query), embedder_name, embedder });
// TODO: would be better to have two distinct functions at this point
let vector_results = search.execute()?;
@ -155,10 +211,10 @@ impl<'a> Search<'a> {
let keyword_results = ScoreWithRatioResult::new(keyword_results, 1.0 - semantic_ratio);
let vector_results = ScoreWithRatioResult::new(vector_results, semantic_ratio);
let merge_results =
let (merge_results, semantic_hit_count) =
ScoreWithRatioResult::merge(vector_results, keyword_results, self.offset, self.limit);
assert!(merge_results.documents_ids.len() <= self.limit);
Ok(merge_results)
Ok((merge_results, Some(semantic_hit_count)))
}
fn results_good_enough(&self, keyword_results: &SearchResult, semantic_ratio: f32) -> bool {

375
milli/src/search/mod.rs
View File

@ -1,25 +1,18 @@
use std::fmt;
use std::ops::ControlFlow;
use std::sync::Arc;
use charabia::normalizer::NormalizerOption;
use charabia::Normalize;
use fst::automaton::{Automaton, Str};
use fst::{IntoStreamer, Streamer};
use levenshtein_automata::{LevenshteinAutomatonBuilder as LevBuilder, DFA};
use once_cell::sync::Lazy;
use roaring::bitmap::RoaringBitmap;
use tracing::error;
pub use self::facet::{FacetDistribution, Filter, OrderBy, DEFAULT_VALUES_PER_FACET};
pub use self::new::matches::{FormatOptions, MatchBounds, MatcherBuilder, MatchingWords};
use self::new::{execute_vector_search, PartialSearchResult};
use crate::error::UserError;
use crate::heed_codec::facet::{FacetGroupKey, FacetGroupValue};
use crate::score_details::{ScoreDetails, ScoringStrategy};
use crate::vector::DistributionShift;
use crate::vector::Embedder;
use crate::{
execute_search, filtered_universe, AscDesc, DefaultSearchLogger, DocumentId, FieldId, Index,
Result, SearchContext,
execute_search, filtered_universe, AscDesc, DefaultSearchLogger, DocumentId, Index, Result,
SearchContext, TimeBudget,
};
// Building these factories is not free.
@ -27,17 +20,20 @@ static LEVDIST0: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(0, true));
static LEVDIST1: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(1, true));
static LEVDIST2: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(2, true));
/// The maximum number of values per facet returned by the facet search route.
const DEFAULT_MAX_NUMBER_OF_VALUES_PER_FACET: usize = 100;
pub mod facet;
mod fst_utils;
pub mod hybrid;
pub mod new;
#[derive(Debug, Clone)]
pub struct SemanticSearch {
vector: Option<Vec<f32>>,
embedder_name: String,
embedder: Arc<Embedder>,
}
pub struct Search<'a> {
query: Option<String>,
vector: Option<Vec<f32>>,
// this should be linked to the String in the query
filter: Option<Filter<'a>>,
offset: usize,
@ -49,18 +45,16 @@ pub struct Search<'a> {
scoring_strategy: ScoringStrategy,
words_limit: usize,
exhaustive_number_hits: bool,
/// TODO: Add semantic ratio or pass it directly to execute_hybrid()
rtxn: &'a heed::RoTxn<'a>,
index: &'a Index,
distribution_shift: Option<DistributionShift>,
embedder_name: Option<String>,
semantic: Option<SemanticSearch>,
time_budget: TimeBudget,
}
impl<'a> Search<'a> {
pub fn new(rtxn: &'a heed::RoTxn, index: &'a Index) -> Search<'a> {
Search {
query: None,
vector: None,
filter: None,
offset: 0,
limit: 20,
@ -73,8 +67,8 @@ impl<'a> Search<'a> {
words_limit: 10,
rtxn,
index,
distribution_shift: None,
embedder_name: None,
semantic: None,
time_budget: TimeBudget::max(),
}
}
@ -83,8 +77,13 @@ impl<'a> Search<'a> {
self
}
pub fn vector(&mut self, vector: Vec<f32>) -> &mut Search<'a> {
self.vector = Some(vector);
pub fn semantic(
&mut self,
embedder_name: String,
embedder: Arc<Embedder>,
vector: Option<Vec<f32>>,
) -> &mut Search<'a> {
self.semantic = Some(SemanticSearch { embedder_name, embedder, vector });
self
}
@ -141,16 +140,8 @@ impl<'a> Search<'a> {
self
}
pub fn distribution_shift(
&mut self,
distribution_shift: Option<DistributionShift>,
) -> &mut Search<'a> {
self.distribution_shift = distribution_shift;
self
}
pub fn embedder_name(&mut self, embedder_name: impl Into<String>) -> &mut Search<'a> {
self.embedder_name = Some(embedder_name.into());
pub fn time_budget(&mut self, time_budget: TimeBudget) -> &mut Search<'a> {
self.time_budget = time_budget;
self
}
@ -164,15 +155,6 @@ impl<'a> Search<'a> {
}
pub fn execute(&self) -> Result<SearchResult> {
let embedder_name;
let embedder_name = match &self.embedder_name {
Some(embedder_name) => embedder_name,
None => {
embedder_name = self.index.default_embedding_name(self.rtxn)?;
&embedder_name
}
};
let mut ctx = SearchContext::new(self.index, self.rtxn);
if let Some(searchable_attributes) = self.searchable_attributes {
@ -180,9 +162,16 @@ impl<'a> Search<'a> {
}
let universe = filtered_universe(&ctx, &self.filter)?;
let PartialSearchResult { located_query_terms, candidates, documents_ids, document_scores } =
match self.vector.as_ref() {
Some(vector) => execute_vector_search(
let PartialSearchResult {
located_query_terms,
candidates,
documents_ids,
document_scores,
degraded,
used_negative_operator,
} = match self.semantic.as_ref() {
Some(SemanticSearch { vector: Some(vector), embedder_name, embedder }) => {
execute_vector_search(
&mut ctx,
vector,
self.scoring_strategy,
@ -191,25 +180,28 @@ impl<'a> Search<'a> {
self.geo_strategy,
self.offset,
self.limit,
self.distribution_shift,
embedder_name,
)?,
None => execute_search(
&mut ctx,
self.query.as_deref(),
self.terms_matching_strategy,
self.scoring_strategy,
self.exhaustive_number_hits,
universe,
&self.sort_criteria,
self.geo_strategy,
self.offset,
self.limit,
Some(self.words_limit),
&mut DefaultSearchLogger,
&mut DefaultSearchLogger,
)?,
};
embedder,
self.time_budget.clone(),
)?
}
_ => execute_search(
&mut ctx,
self.query.as_deref(),
self.terms_matching_strategy,
self.scoring_strategy,
self.exhaustive_number_hits,
universe,
&self.sort_criteria,
self.geo_strategy,
self.offset,
self.limit,
Some(self.words_limit),
&mut DefaultSearchLogger,
&mut DefaultSearchLogger,
self.time_budget.clone(),
)?,
};
// consume context and located_query_terms to build MatchingWords.
let matching_words = match located_query_terms {
@ -217,7 +209,14 @@ impl<'a> Search<'a> {
None => MatchingWords::default(),
};
Ok(SearchResult { matching_words, candidates, document_scores, documents_ids })
Ok(SearchResult {
matching_words,
candidates,
document_scores,
documents_ids,
degraded,
used_negative_operator,
})
}
}
@ -225,7 +224,6 @@ impl fmt::Debug for Search<'_> {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let Search {
query,
vector: _,
filter,
offset,
limit,
@ -238,8 +236,8 @@ impl fmt::Debug for Search<'_> {
exhaustive_number_hits,
rtxn: _,
index: _,
distribution_shift,
embedder_name,
semantic,
time_budget,
} = self;
f.debug_struct("Search")
.field("query", query)
@ -253,8 +251,11 @@ impl fmt::Debug for Search<'_> {
.field("scoring_strategy", scoring_strategy)
.field("exhaustive_number_hits", exhaustive_number_hits)
.field("words_limit", words_limit)
.field("distribution_shift", distribution_shift)
.field("embedder_name", embedder_name)
.field(
"semantic.embedder_name",
&semantic.as_ref().map(|semantic| &semantic.embedder_name),
)
.field("time_budget", time_budget)
.finish()
}
}
@ -265,6 +266,8 @@ pub struct SearchResult {
pub candidates: RoaringBitmap,
pub documents_ids: Vec<DocumentId>,
pub document_scores: Vec<Vec<ScoreDetails>>,
pub degraded: bool,
pub used_negative_operator: bool,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
@ -302,240 +305,6 @@ pub fn build_dfa(word: &str, typos: u8, is_prefix: bool) -> DFA {
}
}
pub struct SearchForFacetValues<'a> {
query: Option<String>,
facet: String,
search_query: Search<'a>,
max_values: usize,
is_hybrid: bool,
}
impl<'a> SearchForFacetValues<'a> {
pub fn new(
facet: String,
search_query: Search<'a>,
is_hybrid: bool,
) -> SearchForFacetValues<'a> {
SearchForFacetValues {
query: None,
facet,
search_query,
max_values: DEFAULT_MAX_NUMBER_OF_VALUES_PER_FACET,
is_hybrid,
}
}
pub fn query(&mut self, query: impl Into<String>) -> &mut Self {
self.query = Some(query.into());
self
}
pub fn max_values(&mut self, max: usize) -> &mut Self {
self.max_values = max;
self
}
fn one_original_value_of(
&self,
field_id: FieldId,
facet_str: &str,
any_docid: DocumentId,
) -> Result<Option<String>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let key: (FieldId, _, &str) = (field_id, any_docid, facet_str);
Ok(index.field_id_docid_facet_strings.get(rtxn, &key)?.map(|v| v.to_owned()))
}
pub fn execute(&self) -> Result<Vec<FacetValueHit>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let filterable_fields = index.filterable_fields(rtxn)?;
if !filterable_fields.contains(&self.facet) {
let (valid_fields, hidden_fields) =
index.remove_hidden_fields(rtxn, filterable_fields)?;
return Err(UserError::InvalidFacetSearchFacetName {
field: self.facet.clone(),
valid_fields,
hidden_fields,
}
.into());
}
let fields_ids_map = index.fields_ids_map(rtxn)?;
let fid = match fields_ids_map.id(&self.facet) {
Some(fid) => fid,
// we return an empty list of results when the attribute has been
// set as filterable but no document contains this field (yet).
None => return Ok(Vec::new()),
};
let fst = match self.search_query.index.facet_id_string_fst.get(rtxn, &fid)? {
Some(fst) => fst,
None => return Ok(vec![]),
};
let search_candidates = self
.search_query
.execute_for_candidates(self.is_hybrid || self.search_query.vector.is_some())?;
match self.query.as_ref() {
Some(query) => {
let options = NormalizerOption { lossy: true, ..Default::default() };
let query = query.normalize(&options);
let query = query.as_ref();
let authorize_typos = self.search_query.index.authorize_typos(rtxn)?;
let field_authorizes_typos =
!self.search_query.index.exact_attributes_ids(rtxn)?.contains(&fid);
if authorize_typos && field_authorizes_typos {
let exact_words_fst = self.search_query.index.exact_words(rtxn)?;
if exact_words_fst.map_or(false, |fst| fst.contains(query)) {
let mut results = vec![];
if fst.contains(query) {
self.fetch_original_facets_using_normalized(
fid,
query,
query,
&search_candidates,
&mut results,
)?;
}
Ok(results)
} else {
let one_typo = self.search_query.index.min_word_len_one_typo(rtxn)?;
let two_typos = self.search_query.index.min_word_len_two_typos(rtxn)?;
let is_prefix = true;
let automaton = if query.len() < one_typo as usize {
build_dfa(query, 0, is_prefix)
} else if query.len() < two_typos as usize {
build_dfa(query, 1, is_prefix)
} else {
build_dfa(query, 2, is_prefix)
};
let mut stream = fst.search(automaton).into_stream();
let mut results = vec![];
while let Some(facet_value) = stream.next() {
let value = std::str::from_utf8(facet_value)?;
if self
.fetch_original_facets_using_normalized(
fid,
value,
query,
&search_candidates,
&mut results,
)?
.is_break()
{
break;
}
}
Ok(results)
}
} else {
let automaton = Str::new(query).starts_with();
let mut stream = fst.search(automaton).into_stream();
let mut results = vec![];
while let Some(facet_value) = stream.next() {
let value = std::str::from_utf8(facet_value)?;
if self
.fetch_original_facets_using_normalized(
fid,
value,
query,
&search_candidates,
&mut results,
)?
.is_break()
{
break;
}
}
Ok(results)
}
}
None => {
let mut results = vec![];
let prefix = FacetGroupKey { field_id: fid, level: 0, left_bound: "" };
for result in index.facet_id_string_docids.prefix_iter(rtxn, &prefix)? {
let (FacetGroupKey { left_bound, .. }, FacetGroupValue { bitmap, .. }) =
result?;
let count = search_candidates.intersection_len(&bitmap);
if count != 0 {
let value = self
.one_original_value_of(fid, left_bound, bitmap.min().unwrap())?
.unwrap_or_else(|| left_bound.to_string());
results.push(FacetValueHit { value, count });
}
if results.len() >= self.max_values {
break;
}
}
Ok(results)
}
}
}
fn fetch_original_facets_using_normalized(
&self,
fid: FieldId,
value: &str,
query: &str,
search_candidates: &RoaringBitmap,
results: &mut Vec<FacetValueHit>,
) -> Result<ControlFlow<()>> {
let index = self.search_query.index;
let rtxn = self.search_query.rtxn;
let database = index.facet_id_normalized_string_strings;
let key = (fid, value);
let original_strings = match database.get(rtxn, &key)? {
Some(original_strings) => original_strings,
None => {
error!("the facet value is missing from the facet database: {key:?}");
return Ok(ControlFlow::Continue(()));
}
};
for original in original_strings {
let key = FacetGroupKey { field_id: fid, level: 0, left_bound: original.as_str() };
let docids = match index.facet_id_string_docids.get(rtxn, &key)? {
Some(FacetGroupValue { bitmap, .. }) => bitmap,
None => {
error!("the facet value is missing from the facet database: {key:?}");
return Ok(ControlFlow::Continue(()));
}
};
let count = search_candidates.intersection_len(&docids);
if count != 0 {
let value = self
.one_original_value_of(fid, &original, docids.min().unwrap())?
.unwrap_or_else(|| query.to_string());
results.push(FacetValueHit { value, count });
}
if results.len() >= self.max_values {
return Ok(ControlFlow::Break(()));
}
}
Ok(ControlFlow::Continue(()))
}
}
#[derive(Debug, Clone, serde::Serialize, PartialEq)]
pub struct FacetValueHit {
/// The original facet value
pub value: String,
/// The number of documents associated to this facet
pub count: u64,
}
#[cfg(test)]
mod test {
#[allow(unused_imports)]

37
milli/src/search/new/bucket_sort.rs
View File

@ -5,12 +5,14 @@ use super::ranking_rules::{BoxRankingRule, RankingRuleQueryTrait};
use super::SearchContext;
use crate::score_details::{ScoreDetails, ScoringStrategy};
use crate::search::new::distinct::{apply_distinct_rule, distinct_single_docid, DistinctOutput};
use crate::Result;
use crate::{Result, TimeBudget};
pub struct BucketSortOutput {
pub docids: Vec<u32>,
pub scores: Vec<Vec<ScoreDetails>>,
pub all_candidates: RoaringBitmap,
pub degraded: bool,
}
// TODO: would probably be good to regroup some of these inside of a struct?
@ -25,6 +27,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
length: usize,
scoring_strategy: ScoringStrategy,
logger: &mut dyn SearchLogger<Q>,
time_budget: TimeBudget,
) -> Result<BucketSortOutput> {
logger.initial_query(query);
logger.ranking_rules(&ranking_rules);
@ -41,6 +44,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
docids: vec![],
scores: vec![],
all_candidates: universe.clone(),
degraded: false,
});
}
if ranking_rules.is_empty() {
@ -74,6 +78,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
scores: vec![Default::default(); results.len()],
docids: results,
all_candidates,
degraded: false,
});
} else {
let docids: Vec<u32> = universe.iter().skip(from).take(length).collect();
@ -81,6 +86,7 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
scores: vec![Default::default(); docids.len()],
docids,
all_candidates: universe.clone(),
degraded: false,
});
};
}
@ -154,6 +160,28 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
}
while valid_docids.len() < length {
if time_budget.exceeded() {
loop {
let bucket = std::mem::take(&mut ranking_rule_universes[cur_ranking_rule_index]);
ranking_rule_scores.push(ScoreDetails::Skipped);
maybe_add_to_results!(bucket);
ranking_rule_scores.pop();
if cur_ranking_rule_index == 0 {
break;
}
back!();
}
return Ok(BucketSortOutput {
scores: valid_scores,
docids: valid_docids,
all_candidates,
degraded: true,
});
}
// The universe for this bucket is zero, so we don't need to sort
// anything, just go back to the parent ranking rule.
if ranking_rule_universes[cur_ranking_rule_index].is_empty()
@ -219,7 +247,12 @@ pub fn bucket_sort<'ctx, Q: RankingRuleQueryTrait>(
)?;
}
Ok(BucketSortOutput { docids: valid_docids, scores: valid_scores, all_candidates })
Ok(BucketSortOutput {
docids: valid_docids,
scores: valid_scores,
all_candidates,
degraded: false,
})
}
/// Add the candidates to the results. Take `distinct`, `from`, `length`, and `cur_offset`

4
milli/src/search/new/matches/matching_words.rs
View File

@ -240,6 +240,7 @@ pub(crate) mod tests {
use super::super::super::located_query_terms_from_tokens;
use super::*;
use crate::index::tests::TempIndex;
use crate::search::new::query_term::ExtractedTokens;
pub(crate) fn temp_index_with_documents() -> TempIndex {
let temp_index = TempIndex::new();
@ -261,7 +262,8 @@ pub(crate) mod tests {
let mut builder = TokenizerBuilder::default();
let tokenizer = builder.build();
let tokens = tokenizer.tokenize("split this world");
let query_terms = located_query_terms_from_tokens(&mut ctx, tokens, None).unwrap();
let ExtractedTokens { query_terms, .. } =
located_query_terms_from_tokens(&mut ctx, tokens, None).unwrap();
let matching_words = MatchingWords::new(ctx, query_terms);
assert_eq!(

3
milli/src/search/new/matches/mod.rs
View File

@ -502,7 +502,7 @@ mod tests {
use super::*;
use crate::index::tests::TempIndex;
use crate::{execute_search, filtered_universe, SearchContext};
use crate::{execute_search, filtered_universe, SearchContext, TimeBudget};
impl<'a> MatcherBuilder<'a> {
fn new_test(rtxn: &'a heed::RoTxn, index: &'a TempIndex, query: &str) -> Self {
@ -522,6 +522,7 @@ mod tests {
Some(10),
&mut crate::DefaultSearchLogger,
&mut crate::DefaultSearchLogger,
TimeBudget::max(),
)
.unwrap();

75
milli/src/search/new/mod.rs
View File

@ -33,7 +33,9 @@ use interner::{DedupInterner, Interner};
pub use logger::visual::VisualSearchLogger;
pub use logger::{DefaultSearchLogger, SearchLogger};
use query_graph::{QueryGraph, QueryNode};
use query_term::{located_query_terms_from_tokens, LocatedQueryTerm, Phrase, QueryTerm};
use query_term::{
located_query_terms_from_tokens, ExtractedTokens, LocatedQueryTerm, Phrase, QueryTerm,
};
use ranking_rules::{
BoxRankingRule, PlaceholderQuery, RankingRule, RankingRuleOutput, RankingRuleQueryTrait,
};
@ -50,9 +52,10 @@ use self::vector_sort::VectorSort;
use crate::error::FieldIdMapMissingEntry;
use crate::score_details::{ScoreDetails, ScoringStrategy};
use crate::search::new::distinct::apply_distinct_rule;
use crate::vector::DistributionShift;
use crate::vector::Embedder;
use crate::{
AscDesc, DocumentId, FieldId, Filter, Index, Member, Result, TermsMatchingStrategy, UserError,
AscDesc, DocumentId, FieldId, Filter, Index, Member, Result, TermsMatchingStrategy, TimeBudget,
UserError,
};
/// A structure used throughout the execution of a search query.
@ -208,6 +211,35 @@ fn resolve_universe(
)
}
#[tracing::instrument(level = "trace", skip_all, target = "search")]
fn resolve_negative_words(
ctx: &mut SearchContext,
negative_words: &[Word],
) -> Result<RoaringBitmap> {
let mut negative_bitmap = RoaringBitmap::new();
for &word in negative_words {
if let Some(bitmap) = ctx.word_docids(word)? {
negative_bitmap |= bitmap;
}
}
Ok(negative_bitmap)
}
#[tracing::instrument(level = "trace", skip_all, target = "search")]
fn resolve_negative_phrases(
ctx: &mut SearchContext,
negative_phrases: &[LocatedQueryTerm],
) -> Result<RoaringBitmap> {
let mut negative_bitmap = RoaringBitmap::new();
for term in negative_phrases {
let query_term = ctx.term_interner.get(term.value);
if let Some(phrase) = query_term.original_phrase() {
negative_bitmap |= ctx.get_phrase_docids(phrase)?;
}
}
Ok(negative_bitmap)
}
/// Return the list of initialised ranking rules to be used for a placeholder search.
fn get_ranking_rules_for_placeholder_search<'ctx>(
ctx: &SearchContext<'ctx>,
@ -266,8 +298,8 @@ fn get_ranking_rules_for_vector<'ctx>(
geo_strategy: geo_sort::Strategy,
limit_plus_offset: usize,
target: &[f32],
distribution_shift: Option<DistributionShift>,
embedder_name: &str,
embedder: &Embedder,
) -> Result<Vec<BoxRankingRule<'ctx, PlaceholderQuery>>> {
// query graph search
@ -293,8 +325,8 @@ fn get_ranking_rules_for_vector<'ctx>(
target.to_vec(),
vector_candidates,
limit_plus_offset,
distribution_shift,
embedder_name,
embedder,
)?;
ranking_rules.push(Box::new(vector_sort));
vector = true;
@ -516,8 +548,9 @@ pub fn execute_vector_search(
geo_strategy: geo_sort::Strategy,
from: usize,
length: usize,
distribution_shift: Option<DistributionShift>,
embedder_name: &str,
embedder: &Embedder,
time_budget: TimeBudget,
) -> Result<PartialSearchResult> {
check_sort_criteria(ctx, sort_criteria.as_ref())?;
@ -529,15 +562,15 @@ pub fn execute_vector_search(
geo_strategy,
from + length,
vector,
distribution_shift,
embedder_name,
embedder,
)?;
let mut placeholder_search_logger = logger::DefaultSearchLogger;
let placeholder_search_logger: &mut dyn SearchLogger<PlaceholderQuery> =
&mut placeholder_search_logger;
let BucketSortOutput { docids, scores, all_candidates } = bucket_sort(
let BucketSortOutput { docids, scores, all_candidates, degraded } = bucket_sort(
ctx,
ranking_rules,
&PlaceholderQuery,
@ -546,6 +579,7 @@ pub fn execute_vector_search(
length,
scoring_strategy,
placeholder_search_logger,
time_budget,
)?;
Ok(PartialSearchResult {
@ -553,6 +587,8 @@ pub fn execute_vector_search(
document_scores: scores,
documents_ids: docids,
located_query_terms: None,
degraded,
used_negative_operator: false,
})
}
@ -572,9 +608,11 @@ pub fn execute_search(
words_limit: Option<usize>,
placeholder_search_logger: &mut dyn SearchLogger<PlaceholderQuery>,
query_graph_logger: &mut dyn SearchLogger<QueryGraph>,
time_budget: TimeBudget,
) -> Result<PartialSearchResult> {
check_sort_criteria(ctx, sort_criteria.as_ref())?;
let mut used_negative_operator = false;
let mut located_query_terms = None;
let query_terms = if let Some(query) = query {
let span = tracing::trace_span!(target: "search::tokens", "tokenizer_builder");
@ -615,7 +653,16 @@ pub fn execute_search(
let tokens = tokenizer.tokenize(query);
drop(entered);
let query_terms = located_query_terms_from_tokens(ctx, tokens, words_limit)?;
let ExtractedTokens { query_terms, negative_words, negative_phrases } =
located_query_terms_from_tokens(ctx, tokens, words_limit)?;
used_negative_operator = !negative_words.is_empty() || !negative_phrases.is_empty();
let ignored_documents = resolve_negative_words(ctx, &negative_words)?;
let ignored_phrases = resolve_negative_phrases(ctx, &negative_phrases)?;
universe -= ignored_documents;
universe -= ignored_phrases;
if query_terms.is_empty() {
// Do a placeholder search instead
None
@ -625,6 +672,7 @@ pub fn execute_search(
} else {
None
};
let bucket_sort_output = if let Some(query_terms) = query_terms {
let (graph, new_located_query_terms) = QueryGraph::from_query(ctx, &query_terms)?;
located_query_terms = Some(new_located_query_terms);
@ -648,6 +696,7 @@ pub fn execute_search(
length,
scoring_strategy,
query_graph_logger,
time_budget,
)?
} else {
let ranking_rules =
@ -661,10 +710,11 @@ pub fn execute_search(
length,
scoring_strategy,
placeholder_search_logger,
time_budget,
)?
};
let BucketSortOutput { docids, scores, mut all_candidates } = bucket_sort_output;
let BucketSortOutput { docids, scores, mut all_candidates, degraded } = bucket_sort_output;
let fields_ids_map = ctx.index.fields_ids_map(ctx.txn)?;
// The candidates is the universe unless the exhaustive number of hits
@ -682,6 +732,8 @@ pub fn execute_search(
document_scores: scores,
documents_ids: docids,
located_query_terms,
degraded,
used_negative_operator,
})
}
@ -742,4 +794,7 @@ pub struct PartialSearchResult {
pub candidates: RoaringBitmap,
pub documents_ids: Vec<DocumentId>,
pub document_scores: Vec<Vec<ScoreDetails>>,
pub degraded: bool,
pub used_negative_operator: bool,
}

9
milli/src/search/new/query_term/mod.rs
View File

@ -9,7 +9,9 @@ use std::ops::RangeInclusive;
use either::Either;
pub use ntypo_subset::NTypoTermSubset;
pub use parse_query::{located_query_terms_from_tokens, make_ngram, number_of_typos_allowed};
pub use parse_query::{
located_query_terms_from_tokens, make_ngram, number_of_typos_allowed, ExtractedTokens,
};
pub use phrase::Phrase;
use super::interner::{DedupInterner, Interned};
@ -478,6 +480,11 @@ impl QueryTerm {
pub fn original_word(&self, ctx: &SearchContext) -> String {
ctx.word_interner.get(self.original).clone()
}
pub fn original_phrase(&self) -> Option<Interned<Phrase>> {
self.zero_typo.phrase
}
pub fn all_computed_derivations(&self) -> (Vec<Interned<String>>, Vec<Interned<Phrase>>) {
let mut words = BTreeSet::new();
let mut phrases = BTreeSet::new();

83
milli/src/search/new/query_term/parse_query.rs
View File

@ -6,20 +6,37 @@ use charabia::{SeparatorKind, TokenKind};
use super::compute_derivations::partially_initialized_term_from_word;
use super::{LocatedQueryTerm, ZeroTypoTerm};
use crate::search::new::query_term::{Lazy, Phrase, QueryTerm};
use crate::search::new::Word;
use crate::{Result, SearchContext, MAX_WORD_LENGTH};
#[derive(Clone)]
/// Extraction of the content of a query.
pub struct ExtractedTokens {
/// The terms to search for in the database.
pub query_terms: Vec<LocatedQueryTerm>,
/// The words that must not appear in the results.
pub negative_words: Vec<Word>,
/// The phrases that must not appear in the results.
pub negative_phrases: Vec<LocatedQueryTerm>,
}
/// Convert the tokenised search query into a list of located query terms.
#[tracing::instrument(level = "trace", skip_all, target = "search::query")]
pub fn located_query_terms_from_tokens(
ctx: &mut SearchContext,
query: NormalizedTokenIter,
words_limit: Option<usize>,
) -> Result<Vec<LocatedQueryTerm>> {
) -> Result<ExtractedTokens> {
let nbr_typos = number_of_typos_allowed(ctx)?;
let mut located_terms = Vec::new();
let mut query_terms = Vec::new();
let mut negative_phrase = false;
let mut phrase: Option<PhraseBuilder> = None;
let mut encountered_whitespace = true;
let mut negative_next_token = false;
let mut negative_words = Vec::new();
let mut negative_phrases = Vec::new();
let parts_limit = words_limit.unwrap_or(usize::MAX);
@ -31,9 +48,10 @@ pub fn located_query_terms_from_tokens(
if token.lemma().is_empty() {
continue;
}
// early return if word limit is exceeded
if located_terms.len() >= parts_limit {
return Ok(located_terms);
if query_terms.len() >= parts_limit {
return Ok(ExtractedTokens { query_terms, negative_words, negative_phrases });
}
match token.kind {
@ -46,6 +64,11 @@ pub fn located_query_terms_from_tokens(
// 3. if the word is the last token of the query we push it as a prefix word.
if let Some(phrase) = &mut phrase {
phrase.push_word(ctx, &token, position)
} else if negative_next_token {
let word = token.lemma().to_string();
let word = Word::Original(ctx.word_interner.insert(word));
negative_words.push(word);
negative_next_token = false;
} else if peekable.peek().is_some() {
match token.kind {
TokenKind::Word => {
@ -61,9 +84,9 @@ pub fn located_query_terms_from_tokens(
value: ctx.term_interner.push(term),
positions: position..=position,
};
located_terms.push(located_term);
query_terms.push(located_term);
}
TokenKind::StopWord | TokenKind::Separator(_) | TokenKind::Unknown => {}
TokenKind::StopWord | TokenKind::Separator(_) | TokenKind::Unknown => (),
}
} else {
let word = token.lemma();
@ -78,7 +101,7 @@ pub fn located_query_terms_from_tokens(
value: ctx.term_interner.push(term),
positions: position..=position,
};
located_terms.push(located_term);
query_terms.push(located_term);
}
}
TokenKind::Separator(separator_kind) => {
@ -94,7 +117,14 @@ pub fn located_query_terms_from_tokens(
let phrase = if separator_kind == SeparatorKind::Hard {
if let Some(phrase) = phrase {
if let Some(located_query_term) = phrase.build(ctx) {
located_terms.push(located_query_term)
// as we are evaluating a negative operator we put the phrase
// in the negative one *but* we don't reset the negative operator
// as we are immediatly starting a new negative phrase.
if negative_phrase {
negative_phrases.push(located_query_term);
} else {
query_terms.push(located_query_term);
}
}
Some(PhraseBuilder::empty())
} else {
@ -115,26 +145,49 @@ pub fn located_query_terms_from_tokens(
// Per the check above, quote_count > 0
quote_count -= 1;
if let Some(located_query_term) = phrase.build(ctx) {
located_terms.push(located_query_term)
// we were evaluating a negative operator so we
// put the phrase in the negative phrases
if negative_phrase {
negative_phrases.push(located_query_term);
negative_phrase = false;
} else {
query_terms.push(located_query_term);
}
}
}
// Start new phrase if the token ends with an opening quote
(quote_count % 2 == 1).then_some(PhraseBuilder::empty())
if quote_count % 2 == 1 {
negative_phrase = negative_next_token;
Some(PhraseBuilder::empty())
} else {
None
}
};
negative_next_token =
phrase.is_none() && token.lemma() == "-" && encountered_whitespace;
}
_ => (),
}
encountered_whitespace =
token.lemma().chars().last().filter(|c| c.is_whitespace()).is_some();
}
// If a quote is never closed, we consider all of the end of the query as a phrase.
if let Some(phrase) = phrase.take() {
if let Some(located_query_term) = phrase.build(ctx) {
located_terms.push(located_query_term);
// put the phrase in the negative set if we are evaluating a negative operator.
if negative_phrase {
negative_phrases.push(located_query_term);
} else {
query_terms.push(located_query_term);
}
}
}
Ok(located_terms)
Ok(ExtractedTokens { query_terms, negative_words, negative_phrases })
}
pub fn number_of_typos_allowed<'ctx>(
@ -315,8 +368,10 @@ mod tests {
let rtxn = index.read_txn()?;
let mut ctx = SearchContext::new(&index, &rtxn);
// panics with `attempt to add with overflow` before <https://github.com/meilisearch/meilisearch/issues/3785>
let located_query_terms = located_query_terms_from_tokens(&mut ctx, tokens, None)?;
assert!(located_query_terms.is_empty());
let ExtractedTokens { query_terms, .. } =
located_query_terms_from_tokens(&mut ctx, tokens, None)?;
assert!(query_terms.is_empty());
Ok(())
}
}

429
milli/src/search/new/tests/cutoff.rs Normal file
View File

@ -0,0 +1,429 @@
//! This module test the search cutoff and ensure a few things:
//! 1. A basic test works and mark the search as degraded
//! 2. A test that ensure the filters are affectively applied even with a cutoff of 0
//! 3. A test that ensure the cutoff works well with the ranking scores
use std::time::Duration;
use big_s::S;
use maplit::hashset;
use meili_snap::snapshot;
use crate::index::tests::TempIndex;
use crate::score_details::{ScoreDetails, ScoringStrategy};
use crate::{Criterion, Filter, Search, TimeBudget};
fn create_index() -> TempIndex {
let index = TempIndex::new();
index
.update_settings(|s| {
s.set_primary_key("id".to_owned());
s.set_searchable_fields(vec!["text".to_owned()]);
s.set_filterable_fields(hashset! { S("id") });
s.set_criteria(vec![Criterion::Words, Criterion::Typo]);
})
.unwrap();
// reverse the ID / insertion order so we see better what was sorted from what got the insertion order ordering
index
.add_documents(documents!([
{
"id": 4,
"text": "hella puppo kefir",
},
{
"id": 3,
"text": "hella puppy kefir",
},
{
"id": 2,
"text": "hello",
},
{
"id": 1,
"text": "hello puppy",
},
{
"id": 0,
"text": "hello puppy kefir",
},
]))
.unwrap();
index
}
#[test]
fn basic_degraded_search() {
let index = create_index();
let rtxn = index.read_txn().unwrap();
let mut search = Search::new(&rtxn, &index);
search.query("hello puppy kefir");
search.limit(3);
search.time_budget(TimeBudget::new(Duration::from_millis(0)));
let result = search.execute().unwrap();
assert!(result.degraded);
}
#[test]
fn degraded_search_cannot_skip_filter() {
let index = create_index();
let rtxn = index.read_txn().unwrap();
let mut search = Search::new(&rtxn, &index);
search.query("hello puppy kefir");
search.limit(100);
search.time_budget(TimeBudget::new(Duration::from_millis(0)));
let filter_condition = Filter::from_str("id > 2").unwrap().unwrap();
search.filter(filter_condition);
let result = search.execute().unwrap();
assert!(result.degraded);
snapshot!(format!("{:?}\n{:?}", result.candidates, result.documents_ids), @r###"
RoaringBitmap<[0, 1]>
[0, 1]
"###);
}
#[test]
#[allow(clippy::format_collect)] // the test is already quite big
fn degraded_search_and_score_details() {
let index = create_index();
let rtxn = index.read_txn().unwrap();
let mut search = Search::new(&rtxn, &index);
search.query("hello puppy kefir");
search.limit(4);
search.scoring_strategy(ScoringStrategy::Detailed);
search.time_budget(TimeBudget::max());
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [4, 1, 0, 3]
Scores: 1.0000 0.9167 0.8333 0.6667
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 1,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 2,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 2,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 2,
},
),
],
]
"###);
// Do ONE loop iteration. Not much can be deduced, almost everyone matched the words first bucket.
search.time_budget(TimeBudget::max().with_stop_after(1));
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [0, 1, 4, 2]
Scores: 0.6667 0.6667 0.6667 0.0000
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Skipped,
],
]
"###);
// Do TWO loop iterations. The first document should be entirely sorted
search.time_budget(TimeBudget::max().with_stop_after(2));
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [4, 0, 1, 2]
Scores: 1.0000 0.6667 0.6667 0.0000
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Skipped,
],
]
"###);
// Do THREE loop iterations. The second document should be entirely sorted as well
search.time_budget(TimeBudget::max().with_stop_after(3));
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [4, 1, 0, 2]
Scores: 1.0000 0.9167 0.6667 0.0000
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 1,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Skipped,
],
[
Skipped,
],
]
"###);
// Do FOUR loop iterations. The third document should be entirely sorted as well
// The words bucket have still not progressed thus the last document doesn't have any info yet.
search.time_budget(TimeBudget::max().with_stop_after(4));
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [4, 1, 0, 2]
Scores: 1.0000 0.9167 0.8333 0.0000
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 1,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 2,
max_typo_count: 3,
},
),
],
[
Skipped,
],
]
"###);
// After SIX loop iteration. The words ranking rule gave us a new bucket.
// Since we reached the limit we were able to early exit without checking the typo ranking rule.
search.time_budget(TimeBudget::max().with_stop_after(6));
let result = search.execute().unwrap();
snapshot!(format!("IDs: {:?}\nScores: {}\nScore Details:\n{:#?}", result.documents_ids, result.document_scores.iter().map(|scores| format!("{:.4} ", ScoreDetails::global_score(scores.iter()))).collect::<String>(), result.document_scores), @r###"
IDs: [4, 1, 0, 3]
Scores: 1.0000 0.9167 0.8333 0.3333
Score Details:
[
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 0,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 1,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 3,
max_matching_words: 3,
},
),
Typo(
Typo {
typo_count: 2,
max_typo_count: 3,
},
),
],
[
Words(
Words {
matching_words: 2,
max_matching_words: 3,
},
),
Skipped,
],
]
"###);
}

1
milli/src/search/new/tests/mod.rs
View File

@ -1,5 +1,6 @@
pub mod attribute_fid;
pub mod attribute_position;
pub mod cutoff;
pub mod distinct;
pub mod exactness;
pub mod geo_sort;

2
milli/src/search/new/tests/typo_proximity.rs
View File

@ -5,7 +5,7 @@ The typo ranking rule should transform the query graph such that it only contain
the combinations of word derivations that it used to compute its bucket.
The proximity ranking rule should then look for proximities only between those specific derivations.
For example, given the the search query `beautiful summer` and the dataset:
For example, given the search query `beautiful summer` and the dataset:
```text
{ "id": 0, "text": "beautigul summer...... beautiful day in the summer" }
{ "id": 1, "text": "beautiful summer" }

34
milli/src/search/new/vector_sort.rs
View File

@ -5,14 +5,14 @@ use roaring::RoaringBitmap;
use super::ranking_rules::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait};
use crate::score_details::{self, ScoreDetails};
use crate::vector::DistributionShift;
use crate::vector::{DistributionShift, Embedder};
use crate::{DocumentId, Result, SearchContext, SearchLogger};
pub struct VectorSort<Q: RankingRuleQueryTrait> {
query: Option<Q>,
target: Vec<f32>,
vector_candidates: RoaringBitmap,
cached_sorted_docids: std::vec::IntoIter<(DocumentId, f32, Vec<f32>)>,
cached_sorted_docids: std::vec::IntoIter<(DocumentId, f32)>,
limit: usize,
distribution_shift: Option<DistributionShift>,
embedder_index: u8,
@ -24,8 +24,8 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
target: Vec<f32>,
vector_candidates: RoaringBitmap,
limit: usize,
distribution_shift: Option<DistributionShift>,
embedder_name: &str,
embedder: &Embedder,
) -> Result<Self> {
let embedder_index = ctx
.index
@ -39,7 +39,7 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
vector_candidates,
cached_sorted_docids: Default::default(),
limit,
distribution_shift,
distribution_shift: embedder.distribution(),
embedder_index,
})
}
@ -70,14 +70,9 @@ impl<Q: RankingRuleQueryTrait> VectorSort<Q> {
for reader in readers.iter() {
let nns_by_vector =
reader.nns_by_vector(ctx.txn, target, self.limit, None, Some(vector_candidates))?;
let vectors: std::result::Result<Vec<_>, _> = nns_by_vector
.iter()
.map(|(docid, _)| reader.item_vector(ctx.txn, *docid).transpose().unwrap())
.collect();
let vectors = vectors?;
results.extend(nns_by_vector.into_iter().zip(vectors).map(|((x, y), z)| (x, y, z)));
results.extend(nns_by_vector.into_iter());
}
results.sort_unstable_by_key(|(_, distance, _)| OrderedFloat(*distance));
results.sort_unstable_by_key(|(_, distance)| OrderedFloat(*distance));
self.cached_sorted_docids = results.into_iter();
Ok(())
@ -118,14 +113,11 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
return Ok(Some(RankingRuleOutput {
query,
candidates: universe.clone(),
score: ScoreDetails::Vector(score_details::Vector {
target_vector: self.target.clone(),
value_similarity: None,
}),
score: ScoreDetails::Vector(score_details::Vector { similarity: None }),
}));
}
for (docid, distance, vector) in self.cached_sorted_docids.by_ref() {
for (docid, distance) in self.cached_sorted_docids.by_ref() {
if vector_candidates.contains(docid) {
let score = 1.0 - distance;
let score = self
@ -135,10 +127,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
return Ok(Some(RankingRuleOutput {
query,
candidates: RoaringBitmap::from_iter([docid]),
score: ScoreDetails::Vector(score_details::Vector {
target_vector: self.target.clone(),
value_similarity: Some((vector, score)),
}),
score: ScoreDetails::Vector(score_details::Vector { similarity: Some(score) }),
}));
}
}
@ -154,10 +143,7 @@ impl<'ctx, Q: RankingRuleQueryTrait> RankingRule<'ctx, Q> for VectorSort<Q> {
return Ok(Some(RankingRuleOutput {
query,
candidates: universe.clone(),
score: ScoreDetails::Vector(score_details::Vector {
target_vector: self.target.clone(),
value_similarity: None,
}),
score: ScoreDetails::Vector(score_details::Vector { similarity: None }),
}));
}

10
milli/src/update/index_documents/extract/extract_vector_points.rs
View File

@ -339,6 +339,7 @@ pub fn extract_embeddings<R: io::Read + io::Seek>(
prompt_reader: grenad::Reader<R>,
indexer: GrenadParameters,
embedder: Arc<Embedder>,
request_threads: &rayon::ThreadPool,
) -> Result<grenad::Reader<BufReader<File>>> {
puffin::profile_function!();
let n_chunks = embedder.chunk_count_hint(); // chunk level parallelism
@ -376,7 +377,10 @@ pub fn extract_embeddings<R: io::Read + io::Seek>(
if chunks.len() == chunks.capacity() {
let chunked_embeds = embedder
.embed_chunks(std::mem::replace(&mut chunks, Vec::with_capacity(n_chunks)))
.embed_chunks(
std::mem::replace(&mut chunks, Vec::with_capacity(n_chunks)),
request_threads,
)
.map_err(crate::vector::Error::from)
.map_err(crate::Error::from)?;
@ -394,7 +398,7 @@ pub fn extract_embeddings<R: io::Read + io::Seek>(
// send last chunk
if !chunks.is_empty() {
let chunked_embeds = embedder
.embed_chunks(std::mem::take(&mut chunks))
.embed_chunks(std::mem::take(&mut chunks), request_threads)
.map_err(crate::vector::Error::from)
.map_err(crate::Error::from)?;
for (docid, embeddings) in chunks_ids
@ -408,7 +412,7 @@ pub fn extract_embeddings<R: io::Read + io::Seek>(
if !current_chunk.is_empty() {
let embeds = embedder
.embed_chunks(vec![std::mem::take(&mut current_chunk)])
.embed_chunks(vec![std::mem::take(&mut current_chunk)], request_threads)
.map_err(crate::vector::Error::from)
.map_err(crate::Error::from)?;

13
milli/src/update/index_documents/extract/mod.rs
View File

@ -238,6 +238,12 @@ fn send_original_documents_data(
let documents_chunk_cloned = original_documents_chunk.clone();
let lmdb_writer_sx_cloned = lmdb_writer_sx.clone();
let request_threads = rayon::ThreadPoolBuilder::new()
.num_threads(crate::vector::REQUEST_PARALLELISM)
.thread_name(|index| format!("embedding-request-{index}"))
.build()?;
rayon::spawn(move || {
for (name, (embedder, prompt)) in embedders {
let result = extract_vector_points(
@ -249,7 +255,12 @@ fn send_original_documents_data(
);
match result {
Ok(ExtractedVectorPoints { manual_vectors, remove_vectors, prompts }) => {
let embeddings = match extract_embeddings(prompts, indexer, embedder.clone()) {
let embeddings = match extract_embeddings(
prompts,
indexer,
embedder.clone(),
&request_threads,
) {
Ok(results) => Some(results),
Err(error) => {
let _ = lmdb_writer_sx_cloned.send(Err(error));

18
milli/src/update/index_documents/mod.rs
View File

@ -2646,6 +2646,13 @@ mod tests {
api_key: Setting::NotSet,
dimensions: Setting::Set(3),
document_template: Setting::NotSet,
url: Setting::NotSet,
query: Setting::NotSet,
input_field: Setting::NotSet,
path_to_embeddings: Setting::NotSet,
embedding_object: Setting::NotSet,
input_type: Setting::NotSet,
distribution: Setting::NotSet,
}),
);
settings.set_embedder_settings(embedders);
@ -2665,7 +2672,16 @@ mod tests {
.unwrap();
let rtxn = index.read_txn().unwrap();
let res = index.search(&rtxn).vector([0.0, 1.0, 2.0].to_vec()).execute().unwrap();
let mut embedding_configs = index.embedding_configs(&rtxn).unwrap();
let (embedder_name, embedder) = embedding_configs.pop().unwrap();
let embedder =
std::sync::Arc::new(crate::vector::Embedder::new(embedder.embedder_options).unwrap());
assert_eq!("manual", embedder_name);
let res = index
.search(&rtxn)
.semantic(embedder_name, embedder, Some([0.0, 1.0, 2.0].to_vec()))
.execute()
.unwrap();
assert_eq!(res.documents_ids.len(), 3);
}

200
milli/src/update/settings.rs
View File

@ -14,12 +14,13 @@ use super::IndexerConfig;
use crate::criterion::Criterion;
use crate::error::UserError;
use crate::index::{DEFAULT_MIN_WORD_LEN_ONE_TYPO, DEFAULT_MIN_WORD_LEN_TWO_TYPOS};
use crate::order_by_map::OrderByMap;
use crate::proximity::ProximityPrecision;
use crate::update::index_documents::IndexDocumentsMethod;
use crate::update::{IndexDocuments, UpdateIndexingStep};
use crate::vector::settings::{check_set, check_unset, EmbedderSource, EmbeddingSettings};
use crate::vector::{Embedder, EmbeddingConfig, EmbeddingConfigs};
use crate::{FieldsIdsMap, Index, OrderBy, Result};
use crate::{FieldsIdsMap, Index, Result};
#[derive(Debug, Clone, PartialEq, Eq, Copy)]
pub enum Setting<T> {
@ -145,10 +146,11 @@ pub struct Settings<'a, 't, 'i> {
/// Attributes on which typo tolerance is disabled.
exact_attributes: Setting<HashSet<String>>,
max_values_per_facet: Setting<usize>,
sort_facet_values_by: Setting<HashMap<String, OrderBy>>,
sort_facet_values_by: Setting<OrderByMap>,
pagination_max_total_hits: Setting<usize>,
proximity_precision: Setting<ProximityPrecision>,
embedder_settings: Setting<BTreeMap<String, Setting<EmbeddingSettings>>>,
search_cutoff: Setting<u64>,
}
impl<'a, 't, 'i> Settings<'a, 't, 'i> {
@ -182,6 +184,7 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
pagination_max_total_hits: Setting::NotSet,
proximity_precision: Setting::NotSet,
embedder_settings: Setting::NotSet,
search_cutoff: Setting::NotSet,
indexer_config,
}
}
@ -340,7 +343,7 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
self.max_values_per_facet = Setting::Reset;
}
pub fn set_sort_facet_values_by(&mut self, value: HashMap<String, OrderBy>) {
pub fn set_sort_facet_values_by(&mut self, value: OrderByMap) {
self.sort_facet_values_by = Setting::Set(value);
}
@ -372,6 +375,14 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
self.embedder_settings = Setting::Reset;
}
pub fn set_search_cutoff(&mut self, value: u64) {
self.search_cutoff = Setting::Set(value);
}
pub fn reset_search_cutoff(&mut self) {
self.search_cutoff = Setting::Reset;
}
#[tracing::instrument(
level = "trace"
skip(self, progress_callback, should_abort, old_fields_ids_map),
@ -965,7 +976,12 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
match joined {
// updated config
EitherOrBoth::Both((name, mut old), (_, new)) => {
changed |= old.apply(new);
changed |= EmbeddingSettings::apply_and_need_reindex(&mut old, new);
if changed {
tracing::debug!(embedder = name, "need reindex");
} else {
tracing::debug!(embedder = name, "skip reindex");
}
let new = validate_embedding_settings(old, &name)?;
new_configs.insert(name, new);
}
@ -1025,6 +1041,24 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
Ok(update)
}
fn update_search_cutoff(&mut self) -> Result<bool> {
let changed = match self.search_cutoff {
Setting::Set(new) => {
let old = self.index.search_cutoff(self.wtxn)?;
if old == Some(new) {
false
} else {
self.index.put_search_cutoff(self.wtxn, new)?;
true
}
}
Setting::Reset => self.index.delete_search_cutoff(self.wtxn)?,
Setting::NotSet => false,
};
Ok(changed)
}
pub fn execute<FP, FA>(mut self, progress_callback: FP, should_abort: FA) -> Result<()>
where
FP: Fn(UpdateIndexingStep) + Sync,
@ -1073,6 +1107,9 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
// 3. Keep the old vectors but reattempt indexing on a prompt change: only actually changed prompt will need embedding + storage
let embedding_configs_updated = self.update_embedding_configs()?;
// never trigger re-indexing
self.update_search_cutoff()?;
if stop_words_updated
|| non_separator_tokens_updated
|| separator_tokens_updated
@ -1131,6 +1168,13 @@ fn validate_prompt(
api_key,
dimensions,
document_template: Setting::Set(template),
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
}) => {
// validate
let template = crate::prompt::Prompt::new(template)
@ -1144,6 +1188,13 @@ fn validate_prompt(
api_key,
dimensions,
document_template: Setting::Set(template),
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
}))
}
new => Ok(new),
@ -1156,8 +1207,21 @@ pub fn validate_embedding_settings(
) -> Result<Setting<EmbeddingSettings>> {
let settings = validate_prompt(name, settings)?;
let Setting::Set(settings) = settings else { return Ok(settings) };
let EmbeddingSettings { source, model, revision, api_key, dimensions, document_template } =
settings;
let EmbeddingSettings {
source,
model,
revision,
api_key,
dimensions,
document_template,
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
} = settings;
if let Some(0) = dimensions.set() {
return Err(crate::error::UserError::InvalidSettingsDimensions {
@ -1166,6 +1230,14 @@ pub fn validate_embedding_settings(
.into());
}
if let Some(url) = url.as_ref().set() {
url::Url::parse(url).map_err(|error| crate::error::UserError::InvalidUrl {
embedder_name: name.to_owned(),
inner_error: error,
url: url.to_owned(),
})?;
}
let Some(inferred_source) = source.set() else {
return Ok(Setting::Set(EmbeddingSettings {
source,
@ -1174,11 +1246,36 @@ pub fn validate_embedding_settings(
api_key,
dimensions,
document_template,
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
}));
};
match inferred_source {
EmbedderSource::OpenAi => {
check_unset(&revision, "revision", inferred_source, name)?;
check_unset(&revision, EmbeddingSettings::REVISION, inferred_source, name)?;
check_unset(&url, EmbeddingSettings::URL, inferred_source, name)?;
check_unset(&query, EmbeddingSettings::QUERY, inferred_source, name)?;
check_unset(&input_field, EmbeddingSettings::INPUT_FIELD, inferred_source, name)?;
check_unset(
&path_to_embeddings,
EmbeddingSettings::PATH_TO_EMBEDDINGS,
inferred_source,
name,
)?;
check_unset(
&embedding_object,
EmbeddingSettings::EMBEDDING_OBJECT,
inferred_source,
name,
)?;
check_unset(&input_type, EmbeddingSettings::INPUT_TYPE, inferred_source, name)?;
if let Setting::Set(model) = &model {
let model = crate::vector::openai::EmbeddingModel::from_name(model.as_str())
.ok_or(crate::error::UserError::InvalidOpenAiModel {
@ -1209,16 +1306,82 @@ pub fn validate_embedding_settings(
}
}
}
EmbedderSource::Ollama => {
// Dimensions get inferred, only model name is required
check_unset(&dimensions, EmbeddingSettings::DIMENSIONS, inferred_source, name)?;
check_set(&model, EmbeddingSettings::MODEL, inferred_source, name)?;
check_unset(&revision, EmbeddingSettings::REVISION, inferred_source, name)?;
check_unset(&query, EmbeddingSettings::QUERY, inferred_source, name)?;
check_unset(&input_field, EmbeddingSettings::INPUT_FIELD, inferred_source, name)?;
check_unset(
&path_to_embeddings,
EmbeddingSettings::PATH_TO_EMBEDDINGS,
inferred_source,
name,
)?;
check_unset(
&embedding_object,
EmbeddingSettings::EMBEDDING_OBJECT,
inferred_source,
name,
)?;
check_unset(&input_type, EmbeddingSettings::INPUT_TYPE, inferred_source, name)?;
}
EmbedderSource::HuggingFace => {
check_unset(&api_key, "apiKey", inferred_source, name)?;
check_unset(&dimensions, "dimensions", inferred_source, name)?;
check_unset(&api_key, EmbeddingSettings::API_KEY, inferred_source, name)?;
check_unset(&dimensions, EmbeddingSettings::DIMENSIONS, inferred_source, name)?;
check_unset(&url, EmbeddingSettings::URL, inferred_source, name)?;
check_unset(&query, EmbeddingSettings::QUERY, inferred_source, name)?;
check_unset(&input_field, EmbeddingSettings::INPUT_FIELD, inferred_source, name)?;
check_unset(
&path_to_embeddings,
EmbeddingSettings::PATH_TO_EMBEDDINGS,
inferred_source,
name,
)?;
check_unset(
&embedding_object,
EmbeddingSettings::EMBEDDING_OBJECT,
inferred_source,
name,
)?;
check_unset(&input_type, EmbeddingSettings::INPUT_TYPE, inferred_source, name)?;
}
EmbedderSource::UserProvided => {
check_unset(&model, "model", inferred_source, name)?;
check_unset(&revision, "revision", inferred_source, name)?;
check_unset(&api_key, "apiKey", inferred_source, name)?;
check_unset(&document_template, "documentTemplate", inferred_source, name)?;
check_set(&dimensions, "dimensions", inferred_source, name)?;
check_unset(&model, EmbeddingSettings::MODEL, inferred_source, name)?;
check_unset(&revision, EmbeddingSettings::REVISION, inferred_source, name)?;
check_unset(&api_key, EmbeddingSettings::API_KEY, inferred_source, name)?;
check_unset(
&document_template,
EmbeddingSettings::DOCUMENT_TEMPLATE,
inferred_source,
name,
)?;
check_set(&dimensions, EmbeddingSettings::DIMENSIONS, inferred_source, name)?;
check_unset(&url, EmbeddingSettings::URL, inferred_source, name)?;
check_unset(&query, EmbeddingSettings::QUERY, inferred_source, name)?;
check_unset(&input_field, EmbeddingSettings::INPUT_FIELD, inferred_source, name)?;
check_unset(
&path_to_embeddings,
EmbeddingSettings::PATH_TO_EMBEDDINGS,
inferred_source,
name,
)?;
check_unset(
&embedding_object,
EmbeddingSettings::EMBEDDING_OBJECT,
inferred_source,
name,
)?;
check_unset(&input_type, EmbeddingSettings::INPUT_TYPE, inferred_source, name)?;
}
EmbedderSource::Rest => {
check_unset(&model, EmbeddingSettings::MODEL, inferred_source, name)?;
check_unset(&revision, EmbeddingSettings::REVISION, inferred_source, name)?;
check_set(&url, EmbeddingSettings::URL, inferred_source, name)?;
}
}
Ok(Setting::Set(EmbeddingSettings {
@ -1228,6 +1391,13 @@ pub fn validate_embedding_settings(
api_key,
dimensions,
document_template,
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
}))
}
@ -2050,6 +2220,7 @@ mod tests {
pagination_max_total_hits,
proximity_precision,
embedder_settings,
search_cutoff,
} = settings;
assert!(matches!(searchable_fields, Setting::NotSet));
assert!(matches!(displayed_fields, Setting::NotSet));
@ -2073,6 +2244,7 @@ mod tests {
assert!(matches!(pagination_max_total_hits, Setting::NotSet));
assert!(matches!(proximity_precision, Setting::NotSet));
assert!(matches!(embedder_settings, Setting::NotSet));
assert!(matches!(search_cutoff, Setting::NotSet));
})
.unwrap();
}

2
milli/src/update/words_prefixes_fst.rs
View File

@ -20,7 +20,7 @@ impl<'t, 'i> WordsPrefixesFst<'t, 'i> {
/// Set the number of words required to make a prefix be part of the words prefixes
/// database. If a word prefix is supposed to match more than this number of words in the
/// dictionnary, therefore this prefix is added to the words prefixes datastructures.
/// dictionary, therefore this prefix is added to the words prefixes datastructures.
///
/// Default value is 100. This value must be higher than 50 and will be clamped
/// to this bound otherwise.

177
milli/src/vector/error.rs
View File

@ -3,7 +3,6 @@ use std::path::PathBuf;
use hf_hub::api::sync::ApiError;
use crate::error::FaultSource;
use crate::vector::openai::OpenAiError;
#[derive(Debug, thiserror::Error)]
#[error("Error while generating embeddings: {inner}")]
@ -51,26 +50,36 @@ pub enum EmbedErrorKind {
TensorValue(candle_core::Error),
#[error("could not run model: {0}")]
ModelForward(candle_core::Error),
#[error("could not reach OpenAI: {0}")]
OpenAiNetwork(reqwest::Error),
#[error("unexpected response from OpenAI: {0}")]
OpenAiUnexpected(reqwest::Error),
#[error("could not authenticate against OpenAI: {0}")]
OpenAiAuth(OpenAiError),
#[error("sent too many requests to OpenAI: {0}")]
OpenAiTooManyRequests(OpenAiError),
#[error("received internal error from OpenAI: {0:?}")]
OpenAiInternalServerError(Option<OpenAiError>),
#[error("sent too many tokens in a request to OpenAI: {0}")]
OpenAiTooManyTokens(OpenAiError),
#[error("received unhandled HTTP status code {0} from OpenAI")]
OpenAiUnhandledStatusCode(u16),
#[error("attempt to embed the following text in a configuration where embeddings must be user provided: {0:?}")]
ManualEmbed(String),
#[error("could not initialize asynchronous runtime: {0}")]
OpenAiRuntimeInit(std::io::Error),
#[error("initializing web client for sending embedding requests failed: {0}")]
InitWebClient(reqwest::Error),
#[error("model not found. Meilisearch will not automatically download models from the Ollama library, please pull the model manually: {0:?}")]
OllamaModelNotFoundError(Option<String>),
#[error("error deserialization the response body as JSON: {0}")]
RestResponseDeserialization(std::io::Error),
#[error("component `{0}` not found in path `{1}` in response: `{2}`")]
RestResponseMissingEmbeddings(String, String, String),
#[error("unexpected format of the embedding response: {0}")]
RestResponseFormat(serde_json::Error),
#[error("expected a response containing {0} embeddings, got only {1}")]
RestResponseEmbeddingCount(usize, usize),
#[error("could not authenticate against embedding server: {0:?}")]
RestUnauthorized(Option<String>),
#[error("sent too many requests to embedding server: {0:?}")]
RestTooManyRequests(Option<String>),
#[error("sent a bad request to embedding server: {0:?}")]
RestBadRequest(Option<String>),
#[error("received internal error from embedding server: {0:?}")]
RestInternalServerError(u16, Option<String>),
#[error("received HTTP {0} from embedding server: {0:?}")]
RestOtherStatusCode(u16, Option<String>),
#[error("could not reach embedding server: {0}")]
RestNetwork(ureq::Transport),
#[error("was expected '{}' to be an object in query '{0}'", .1.join("."))]
RestNotAnObject(serde_json::Value, Vec<String>),
#[error("while embedding tokenized, was expecting embeddings of dimension `{0}`, got embeddings of dimensions `{1}`")]
OpenAiUnexpectedDimension(usize, usize),
#[error("no embedding was produced")]
MissingEmbedding,
}
impl EmbedError {
@ -90,44 +99,101 @@ impl EmbedError {
Self { kind: EmbedErrorKind::ModelForward(inner), fault: FaultSource::Runtime }
}
pub fn openai_network(inner: reqwest::Error) -> Self {
Self { kind: EmbedErrorKind::OpenAiNetwork(inner), fault: FaultSource::Runtime }
}
pub fn openai_unexpected(inner: reqwest::Error) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiUnexpected(inner), fault: FaultSource::Bug }
}
pub(crate) fn openai_auth_error(inner: OpenAiError) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiAuth(inner), fault: FaultSource::User }
}
pub(crate) fn openai_too_many_requests(inner: OpenAiError) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiTooManyRequests(inner), fault: FaultSource::Runtime }
}
pub(crate) fn openai_internal_server_error(inner: Option<OpenAiError>) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiInternalServerError(inner), fault: FaultSource::Runtime }
}
pub(crate) fn openai_too_many_tokens(inner: OpenAiError) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiTooManyTokens(inner), fault: FaultSource::Bug }
}
pub(crate) fn openai_unhandled_status_code(code: u16) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiUnhandledStatusCode(code), fault: FaultSource::Bug }
}
pub(crate) fn embed_on_manual_embedder(texts: String) -> EmbedError {
Self { kind: EmbedErrorKind::ManualEmbed(texts), fault: FaultSource::User }
}
pub(crate) fn openai_runtime_init(inner: std::io::Error) -> EmbedError {
Self { kind: EmbedErrorKind::OpenAiRuntimeInit(inner), fault: FaultSource::Runtime }
pub(crate) fn ollama_model_not_found(inner: Option<String>) -> EmbedError {
Self { kind: EmbedErrorKind::OllamaModelNotFoundError(inner), fault: FaultSource::User }
}
pub fn openai_initialize_web_client(inner: reqwest::Error) -> Self {
Self { kind: EmbedErrorKind::InitWebClient(inner), fault: FaultSource::Runtime }
pub(crate) fn rest_response_deserialization(error: std::io::Error) -> EmbedError {
Self {
kind: EmbedErrorKind::RestResponseDeserialization(error),
fault: FaultSource::Runtime,
}
}
pub(crate) fn rest_response_missing_embeddings<S: AsRef<str>>(
response: serde_json::Value,
component: &str,
response_field: &[S],
) -> EmbedError {
let response_field: Vec<&str> = response_field.iter().map(AsRef::as_ref).collect();
let response_field = response_field.join(".");
Self {
kind: EmbedErrorKind::RestResponseMissingEmbeddings(
component.to_owned(),
response_field,
serde_json::to_string_pretty(&response).unwrap_or_default(),
),
fault: FaultSource::Undecided,
}
}
pub(crate) fn rest_response_format(error: serde_json::Error) -> EmbedError {
Self { kind: EmbedErrorKind::RestResponseFormat(error), fault: FaultSource::Undecided }
}
pub(crate) fn rest_response_embedding_count(expected: usize, got: usize) -> EmbedError {
Self {
kind: EmbedErrorKind::RestResponseEmbeddingCount(expected, got),
fault: FaultSource::Runtime,
}
}
pub(crate) fn rest_unauthorized(error_response: Option<String>) -> EmbedError {
Self { kind: EmbedErrorKind::RestUnauthorized(error_response), fault: FaultSource::User }
}
pub(crate) fn rest_too_many_requests(error_response: Option<String>) -> EmbedError {
Self {
kind: EmbedErrorKind::RestTooManyRequests(error_response),
fault: FaultSource::Runtime,
}
}
pub(crate) fn rest_bad_request(error_response: Option<String>) -> EmbedError {
Self { kind: EmbedErrorKind::RestBadRequest(error_response), fault: FaultSource::User }
}
pub(crate) fn rest_internal_server_error(
code: u16,
error_response: Option<String>,
) -> EmbedError {
Self {
kind: EmbedErrorKind::RestInternalServerError(code, error_response),
fault: FaultSource::Runtime,
}
}
pub(crate) fn rest_other_status_code(code: u16, error_response: Option<String>) -> EmbedError {
Self {
kind: EmbedErrorKind::RestOtherStatusCode(code, error_response),
fault: FaultSource::Undecided,
}
}
pub(crate) fn rest_network(transport: ureq::Transport) -> EmbedError {
Self { kind: EmbedErrorKind::RestNetwork(transport), fault: FaultSource::Runtime }
}
pub(crate) fn rest_not_an_object(
query: serde_json::Value,
input_path: Vec<String>,
) -> EmbedError {
Self { kind: EmbedErrorKind::RestNotAnObject(query, input_path), fault: FaultSource::User }
}
pub(crate) fn openai_unexpected_dimension(expected: usize, got: usize) -> EmbedError {
Self {
kind: EmbedErrorKind::OpenAiUnexpectedDimension(expected, got),
fault: FaultSource::Runtime,
}
}
pub(crate) fn missing_embedding() -> EmbedError {
Self { kind: EmbedErrorKind::MissingEmbedding, fault: FaultSource::Undecided }
}
}
@ -188,16 +254,12 @@ impl NewEmbedderError {
Self { kind: NewEmbedderErrorKind::LoadModel(inner), fault: FaultSource::Runtime }
}
pub fn hf_could_not_determine_dimension(inner: EmbedError) -> NewEmbedderError {
pub fn could_not_determine_dimension(inner: EmbedError) -> NewEmbedderError {
Self {
kind: NewEmbedderErrorKind::CouldNotDetermineDimension(inner),
fault: FaultSource::Runtime,
}
}
pub fn openai_invalid_api_key_format(inner: reqwest::header::InvalidHeaderValue) -> Self {
Self { kind: NewEmbedderErrorKind::InvalidApiKeyFormat(inner), fault: FaultSource::User }
}
}
#[derive(Debug, thiserror::Error)]
@ -244,7 +306,4 @@ pub enum NewEmbedderErrorKind {
CouldNotDetermineDimension(EmbedError),
#[error("loading model failed: {0}")]
LoadModel(candle_core::Error),
// openai
#[error("The API key passed to Authorization error was in an invalid format: {0}")]
InvalidApiKeyFormat(reqwest::header::InvalidHeaderValue),
}

27
milli/src/vector/hf.rs
View File

@ -33,6 +33,7 @@ enum WeightSource {
pub struct EmbedderOptions {
pub model: String,
pub revision: Option<String>,
pub distribution: Option<DistributionShift>,
}
impl EmbedderOptions {
@ -40,6 +41,7 @@ impl EmbedderOptions {
Self {
model: "BAAI/bge-base-en-v1.5".to_string(),
revision: Some("617ca489d9e86b49b8167676d8220688b99db36e".into()),
distribution: None,
}
}
}
@ -87,11 +89,11 @@ impl Embedder {
let config = api.get("config.json").map_err(NewEmbedderError::api_get)?;
let tokenizer = api.get("tokenizer.json").map_err(NewEmbedderError::api_get)?;
let (weights, source) = {
api.get("pytorch_model.bin")
.map(|filename| (filename, WeightSource::Pytorch))
api.get("model.safetensors")
.map(|filename| (filename, WeightSource::Safetensors))
.or_else(|_| {
api.get("model.safetensors")
.map(|filename| (filename, WeightSource::Safetensors))
api.get("pytorch_model.bin")
.map(|filename| (filename, WeightSource::Pytorch))
})
.map_err(NewEmbedderError::api_get)?
};
@ -131,7 +133,7 @@ impl Embedder {
let embeddings = this
.embed(vec!["test".into()])
.map_err(NewEmbedderError::hf_could_not_determine_dimension)?;
.map_err(NewEmbedderError::could_not_determine_dimension)?;
this.dimensions = embeddings.first().unwrap().dimension();
Ok(this)
@ -193,10 +195,15 @@ impl Embedder {
}
pub fn distribution(&self) -> Option<DistributionShift> {
if self.options.model == "BAAI/bge-base-en-v1.5" {
Some(DistributionShift { current_mean: 0.85, current_sigma: 0.1 })
} else {
None
}
self.options.distribution.or_else(|| {
if self.options.model == "BAAI/bge-base-en-v1.5" {
Some(DistributionShift {
current_mean: ordered_float::OrderedFloat(0.85),
current_sigma: ordered_float::OrderedFloat(0.1),
})
} else {
None
}
})
}
}

10
milli/src/vector/manual.rs
View File

@ -1,19 +1,21 @@
use super::error::EmbedError;
use super::Embeddings;
use super::{DistributionShift, Embeddings};
#[derive(Debug, Clone, Copy)]
pub struct Embedder {
dimensions: usize,
distribution: Option<DistributionShift>,
}
#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
pub struct EmbedderOptions {
pub dimensions: usize,
pub distribution: Option<DistributionShift>,
}
impl Embedder {
pub fn new(options: EmbedderOptions) -> Self {
Self { dimensions: options.dimensions }
Self { dimensions: options.dimensions, distribution: options.distribution }
}
pub fn embed(&self, mut texts: Vec<String>) -> Result<Vec<Embeddings<f32>>, EmbedError> {
@ -31,4 +33,8 @@ impl Embedder {
) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
text_chunks.into_iter().map(|prompts| self.embed(prompts)).collect()
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.distribution
}
}

211
milli/src/vector/mod.rs
View File

@ -1,6 +1,10 @@
use std::collections::HashMap;
use std::sync::Arc;
use deserr::{DeserializeError, Deserr};
use ordered_float::OrderedFloat;
use serde::{Deserialize, Serialize};
use self::error::{EmbedError, NewEmbedderError};
use crate::prompt::{Prompt, PromptData};
@ -10,50 +14,71 @@ pub mod manual;
pub mod openai;
pub mod settings;
pub mod ollama;
pub mod rest;
pub use self::error::Error;
pub type Embedding = Vec<f32>;
pub const REQUEST_PARALLELISM: usize = 40;
/// One or multiple embeddings stored consecutively in a flat vector.
pub struct Embeddings<F> {
data: Vec<F>,
dimension: usize,
}
impl<F> Embeddings<F> {
/// Declares an empty vector of embeddings of the specified dimensions.
pub fn new(dimension: usize) -> Self {
Self { data: Default::default(), dimension }
}
/// Declares a vector of embeddings containing a single element.
///
/// The dimension is inferred from the length of the passed embedding.
pub fn from_single_embedding(embedding: Vec<F>) -> Self {
Self { dimension: embedding.len(), data: embedding }
}
/// Declares a vector of embeddings from its components.
///
/// `data.len()` must be a multiple of `dimension`, otherwise an error is returned.
pub fn from_inner(data: Vec<F>, dimension: usize) -> Result<Self, Vec<F>> {
let mut this = Self::new(dimension);
this.append(data)?;
Ok(this)
}
/// Returns the number of embeddings in this vector of embeddings.
pub fn embedding_count(&self) -> usize {
self.data.len() / self.dimension
}
/// Dimension of a single embedding.
pub fn dimension(&self) -> usize {
self.dimension
}
/// Deconstructs self into the inner flat vector.
pub fn into_inner(self) -> Vec<F> {
self.data
}
/// A reference to the inner flat vector.
pub fn as_inner(&self) -> &[F] {
&self.data
}
/// Iterates over the embeddings contained in the flat vector.
pub fn iter(&self) -> impl Iterator<Item = &'_ [F]> + '_ {
self.data.as_slice().chunks_exact(self.dimension)
}
/// Push an embedding at the end of the embeddings.
///
/// If `embedding.len() != self.dimension`, then the push operation fails.
pub fn push(&mut self, mut embedding: Vec<F>) -> Result<(), Vec<F>> {
if embedding.len() != self.dimension {
return Err(embedding);
@ -62,6 +87,9 @@ impl<F> Embeddings<F> {
Ok(())
}
/// Append a flat vector of embeddings a the end of the embeddings.
///
/// If `embeddings.len() % self.dimension != 0`, then the append operation fails.
pub fn append(&mut self, mut embeddings: Vec<F>) -> Result<(), Vec<F>> {
if embeddings.len() % self.dimension != 0 {
return Err(embeddings);
@ -71,44 +99,68 @@ impl<F> Embeddings<F> {
}
}
/// An embedder can be used to transform text into embeddings.
#[derive(Debug)]
pub enum Embedder {
/// An embedder based on running local models, fetched from the Hugging Face Hub.
HuggingFace(hf::Embedder),
/// An embedder based on making embedding queries against the OpenAI API.
OpenAi(openai::Embedder),
/// An embedder based on the user providing the embeddings in the documents and queries.
UserProvided(manual::Embedder),
/// An embedder based on making embedding queries against an <https://ollama.com> embedding server.
Ollama(ollama::Embedder),
/// An embedder based on making embedding queries against a generic JSON/REST embedding server.
Rest(rest::Embedder),
}
/// Configuration for an embedder.
#[derive(Debug, Clone, Default, serde::Deserialize, serde::Serialize)]
pub struct EmbeddingConfig {
/// Options of the embedder, specific to each kind of embedder
pub embedder_options: EmbedderOptions,
/// Document template
pub prompt: PromptData,
// TODO: add metrics and anything needed
}
/// Map of embedder configurations.
///
/// Each configuration is mapped to a name.
#[derive(Clone, Default)]
pub struct EmbeddingConfigs(HashMap<String, (Arc<Embedder>, Arc<Prompt>)>);
impl EmbeddingConfigs {
/// Create the map from its internal component.s
pub fn new(data: HashMap<String, (Arc<Embedder>, Arc<Prompt>)>) -> Self {
Self(data)
}
/// Get an embedder configuration and template from its name.
pub fn get(&self, name: &str) -> Option<(Arc<Embedder>, Arc<Prompt>)> {
self.0.get(name).cloned()
}
/// Get the default embedder configuration, if any.
pub fn get_default(&self) -> Option<(Arc<Embedder>, Arc<Prompt>)> {
self.get_default_embedder_name().and_then(|default| self.get(&default))
self.get(self.get_default_embedder_name())
}
pub fn get_default_embedder_name(&self) -> Option<String> {
/// Get the name of the default embedder configuration.
///
/// The default embedder is determined as follows:
///
/// - If there is only one embedder, it is always the default.
/// - If there are multiple embedders and one of them is called `default`, then that one is the default embedder.
/// - In all other cases, there is no default embedder.
pub fn get_default_embedder_name(&self) -> &str {
let mut it = self.0.keys();
let first_name = it.next();
let second_name = it.next();
match (first_name, second_name) {
(None, _) => None,
(Some(first), None) => Some(first.to_owned()),
(Some(_), Some(_)) => Some("default".to_owned()),
(None, _) => "default",
(Some(first), None) => first,
(Some(_), Some(_)) => "default",
}
}
}
@ -123,11 +175,14 @@ impl IntoIterator for EmbeddingConfigs {
}
}
/// Options of an embedder, specific to each kind of embedder.
#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
pub enum EmbedderOptions {
HuggingFace(hf::EmbedderOptions),
OpenAi(openai::EmbedderOptions),
Ollama(ollama::EmbedderOptions),
UserProvided(manual::EmbedderOptions),
Rest(rest::EmbedderOptions),
}
impl Default for EmbedderOptions {
@ -137,91 +192,204 @@ impl Default for EmbedderOptions {
}
impl EmbedderOptions {
/// Default options for the Hugging Face embedder
pub fn huggingface() -> Self {
Self::HuggingFace(hf::EmbedderOptions::new())
}
/// Default options for the OpenAI embedder
pub fn openai(api_key: Option<String>) -> Self {
Self::OpenAi(openai::EmbedderOptions::with_default_model(api_key))
}
pub fn ollama(api_key: Option<String>, url: Option<String>) -> Self {
Self::Ollama(ollama::EmbedderOptions::with_default_model(api_key, url))
}
}
impl Embedder {
/// Spawns a new embedder built from its options.
pub fn new(options: EmbedderOptions) -> std::result::Result<Self, NewEmbedderError> {
Ok(match options {
EmbedderOptions::HuggingFace(options) => Self::HuggingFace(hf::Embedder::new(options)?),
EmbedderOptions::OpenAi(options) => Self::OpenAi(openai::Embedder::new(options)?),
EmbedderOptions::Ollama(options) => Self::Ollama(ollama::Embedder::new(options)?),
EmbedderOptions::UserProvided(options) => {
Self::UserProvided(manual::Embedder::new(options))
}
EmbedderOptions::Rest(options) => Self::Rest(rest::Embedder::new(options)?),
})
}
pub async fn embed(
/// Embed one or multiple texts.
///
/// Each text can be embedded as one or multiple embeddings.
pub fn embed(
&self,
texts: Vec<String>,
) -> std::result::Result<Vec<Embeddings<f32>>, EmbedError> {
match self {
Embedder::HuggingFace(embedder) => embedder.embed(texts),
Embedder::OpenAi(embedder) => {
let client = embedder.new_client()?;
embedder.embed(texts, &client).await
}
Embedder::OpenAi(embedder) => embedder.embed(texts),
Embedder::Ollama(embedder) => embedder.embed(texts),
Embedder::UserProvided(embedder) => embedder.embed(texts),
Embedder::Rest(embedder) => embedder.embed(texts),
}
}
/// # Panics
pub fn embed_one(&self, text: String) -> std::result::Result<Embedding, EmbedError> {
let mut embeddings = self.embed(vec![text])?;
let embeddings = embeddings.pop().ok_or_else(EmbedError::missing_embedding)?;
Ok(if embeddings.iter().nth(1).is_some() {
tracing::warn!("Ignoring embeddings past the first one in long search query");
embeddings.iter().next().unwrap().to_vec()
} else {
embeddings.into_inner()
})
}
/// Embed multiple chunks of texts.
///
/// - if called from an asynchronous context
/// Each chunk is composed of one or multiple texts.
pub fn embed_chunks(
&self,
text_chunks: Vec<Vec<String>>,
threads: &rayon::ThreadPool,
) -> std::result::Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
match self {
Embedder::HuggingFace(embedder) => embedder.embed_chunks(text_chunks),
Embedder::OpenAi(embedder) => embedder.embed_chunks(text_chunks),
Embedder::OpenAi(embedder) => embedder.embed_chunks(text_chunks, threads),
Embedder::Ollama(embedder) => embedder.embed_chunks(text_chunks, threads),
Embedder::UserProvided(embedder) => embedder.embed_chunks(text_chunks),
Embedder::Rest(embedder) => embedder.embed_chunks(text_chunks, threads),
}
}
/// Indicates the preferred number of chunks to pass to [`Self::embed_chunks`]
pub fn chunk_count_hint(&self) -> usize {
match self {
Embedder::HuggingFace(embedder) => embedder.chunk_count_hint(),
Embedder::OpenAi(embedder) => embedder.chunk_count_hint(),
Embedder::Ollama(embedder) => embedder.chunk_count_hint(),
Embedder::UserProvided(_) => 1,
Embedder::Rest(embedder) => embedder.chunk_count_hint(),
}
}
/// Indicates the preferred number of texts in a single chunk passed to [`Self::embed`]
pub fn prompt_count_in_chunk_hint(&self) -> usize {
match self {
Embedder::HuggingFace(embedder) => embedder.prompt_count_in_chunk_hint(),
Embedder::OpenAi(embedder) => embedder.prompt_count_in_chunk_hint(),
Embedder::Ollama(embedder) => embedder.prompt_count_in_chunk_hint(),
Embedder::UserProvided(_) => 1,
Embedder::Rest(embedder) => embedder.prompt_count_in_chunk_hint(),
}
}
/// Indicates the dimensions of a single embedding produced by the embedder.
pub fn dimensions(&self) -> usize {
match self {
Embedder::HuggingFace(embedder) => embedder.dimensions(),
Embedder::OpenAi(embedder) => embedder.dimensions(),
Embedder::Ollama(embedder) => embedder.dimensions(),
Embedder::UserProvided(embedder) => embedder.dimensions(),
Embedder::Rest(embedder) => embedder.dimensions(),
}
}
/// An optional distribution used to apply an affine transformation to the similarity score of a document.
pub fn distribution(&self) -> Option<DistributionShift> {
match self {
Embedder::HuggingFace(embedder) => embedder.distribution(),
Embedder::OpenAi(embedder) => embedder.distribution(),
Embedder::UserProvided(_embedder) => None,
Embedder::Ollama(embedder) => embedder.distribution(),
Embedder::UserProvided(embedder) => embedder.distribution(),
Embedder::Rest(embedder) => embedder.distribution(),
}
}
}
#[derive(Debug, Clone, Copy)]
/// Describes the mean and sigma of distribution of embedding similarity in the embedding space.
///
/// The intended use is to make the similarity score more comparable to the regular ranking score.
/// This allows to correct effects where results are too "packed" around a certain value.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Deserialize, Serialize)]
#[serde(from = "DistributionShiftSerializable")]
#[serde(into = "DistributionShiftSerializable")]
pub struct DistributionShift {
pub current_mean: f32,
pub current_sigma: f32,
/// Value where the results are "packed".
///
/// Similarity scores are translated so that they are packed around 0.5 instead
pub current_mean: OrderedFloat<f32>,
/// standard deviation of a similarity score.
///
/// Set below 0.4 to make the results less packed around the mean, and above 0.4 to make them more packed.
pub current_sigma: OrderedFloat<f32>,
}
impl<E> Deserr<E> for DistributionShift
where
E: DeserializeError,
{
fn deserialize_from_value<V: deserr::IntoValue>(
value: deserr::Value<V>,
location: deserr::ValuePointerRef,
) -> Result<Self, E> {
let value = DistributionShiftSerializable::deserialize_from_value(value, location)?;
if value.mean < 0. || value.mean > 1. {
return Err(deserr::take_cf_content(E::error::<std::convert::Infallible>(
None,
deserr::ErrorKind::Unexpected {
msg: format!(
"the distribution mean must be in the range [0, 1], got {}",
value.mean
),
},
location,
)));
}
if value.sigma <= 0. || value.sigma > 1. {
return Err(deserr::take_cf_content(E::error::<std::convert::Infallible>(
None,
deserr::ErrorKind::Unexpected {
msg: format!(
"the distribution sigma must be in the range ]0, 1], got {}",
value.sigma
),
},
location,
)));
}
Ok(value.into())
}
}
#[derive(Serialize, Deserialize, Deserr)]
#[serde(deny_unknown_fields)]
#[deserr(deny_unknown_fields)]
struct DistributionShiftSerializable {
mean: f32,
sigma: f32,
}
impl From<DistributionShift> for DistributionShiftSerializable {
fn from(
DistributionShift {
current_mean: OrderedFloat(current_mean),
current_sigma: OrderedFloat(current_sigma),
}: DistributionShift,
) -> Self {
Self { mean: current_mean, sigma: current_sigma }
}
}
impl From<DistributionShiftSerializable> for DistributionShift {
fn from(DistributionShiftSerializable { mean, sigma }: DistributionShiftSerializable) -> Self {
Self { current_mean: OrderedFloat(mean), current_sigma: OrderedFloat(sigma) }
}
}
impl DistributionShift {
@ -230,11 +398,13 @@ impl DistributionShift {
if sigma <= 0.0 {
None
} else {
Some(Self { current_mean: mean, current_sigma: sigma })
Some(Self { current_mean: OrderedFloat(mean), current_sigma: OrderedFloat(sigma) })
}
}
pub fn shift(&self, score: f32) -> f32 {
let current_mean = self.current_mean.0;
let current_sigma = self.current_sigma.0;
// <https://math.stackexchange.com/a/2894689>
// We're somewhat abusively mapping the distribution of distances to a gaussian.
// The parameters we're given is the mean and sigma of the native result distribution.
@ -244,9 +414,9 @@ impl DistributionShift {
let target_sigma = 0.4;
// a^2 sig1^2 = sig2^2 => a^2 = sig2^2 / sig1^2 => a = sig2 / sig1, assuming a, sig1, and sig2 positive.
let factor = target_sigma / self.current_sigma;
let factor = target_sigma / current_sigma;
// a*mu1 + b = mu2 => b = mu2 - a*mu1
let offset = target_mean - (factor * self.current_mean);
let offset = target_mean - (factor * current_mean);
let mut score = factor * score + offset;
@ -262,6 +432,7 @@ impl DistributionShift {
}
}
/// Whether CUDA is supported in this version of Meilisearch.
pub const fn is_cuda_enabled() -> bool {
cfg!(feature = "cuda")
}

101
milli/src/vector/ollama.rs Normal file
View File

@ -0,0 +1,101 @@
use rayon::iter::{IntoParallelIterator as _, ParallelIterator as _};
use super::error::{EmbedError, EmbedErrorKind, NewEmbedderError, NewEmbedderErrorKind};
use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
use super::{DistributionShift, Embeddings};
#[derive(Debug)]
pub struct Embedder {
rest_embedder: RestEmbedder,
}
#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
pub struct EmbedderOptions {
pub embedding_model: String,
pub url: Option<String>,
pub api_key: Option<String>,
pub distribution: Option<DistributionShift>,
}
impl EmbedderOptions {
pub fn with_default_model(api_key: Option<String>, url: Option<String>) -> Self {
Self { embedding_model: "nomic-embed-text".into(), api_key, url, distribution: None }
}
}
impl Embedder {
pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
let model = options.embedding_model.as_str();
let rest_embedder = match RestEmbedder::new(RestEmbedderOptions {
api_key: options.api_key,
dimensions: None,
distribution: options.distribution,
url: options.url.unwrap_or_else(get_ollama_path),
query: serde_json::json!({
"model": model,
}),
input_field: vec!["prompt".to_owned()],
path_to_embeddings: Default::default(),
embedding_object: vec!["embedding".to_owned()],
input_type: super::rest::InputType::Text,
}) {
Ok(embedder) => embedder,
Err(NewEmbedderError {
kind:
NewEmbedderErrorKind::CouldNotDetermineDimension(EmbedError {
kind: super::error::EmbedErrorKind::RestOtherStatusCode(404, error),
fault: _,
}),
fault: _,
}) => {
return Err(NewEmbedderError::could_not_determine_dimension(
EmbedError::ollama_model_not_found(error),
))
}
Err(error) => return Err(error),
};
Ok(Self { rest_embedder })
}
pub fn embed(&self, texts: Vec<String>) -> Result<Vec<Embeddings<f32>>, EmbedError> {
match self.rest_embedder.embed(texts) {
Ok(embeddings) => Ok(embeddings),
Err(EmbedError { kind: EmbedErrorKind::RestOtherStatusCode(404, error), fault: _ }) => {
Err(EmbedError::ollama_model_not_found(error))
}
Err(error) => Err(error),
}
}
pub fn embed_chunks(
&self,
text_chunks: Vec<Vec<String>>,
threads: &rayon::ThreadPool,
) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
threads.install(move || {
text_chunks.into_par_iter().map(move |chunk| self.embed(chunk)).collect()
})
}
pub fn chunk_count_hint(&self) -> usize {
self.rest_embedder.chunk_count_hint()
}
pub fn prompt_count_in_chunk_hint(&self) -> usize {
self.rest_embedder.prompt_count_in_chunk_hint()
}
pub fn dimensions(&self) -> usize {
self.rest_embedder.dimensions()
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.rest_embedder.distribution()
}
}
fn get_ollama_path() -> String {
// Important: Hostname not enough, has to be entire path to embeddings endpoint
std::env::var("MEILI_OLLAMA_URL").unwrap_or("http://localhost:11434/api/embeddings".to_string())
}

574
milli/src/vector/openai.rs
View File

@ -1,23 +1,47 @@
use std::fmt::Display;
use reqwest::StatusCode;
use serde::{Deserialize, Serialize};
use ordered_float::OrderedFloat;
use rayon::iter::{IntoParallelIterator, ParallelIterator as _};
use super::error::{EmbedError, NewEmbedderError};
use super::{DistributionShift, Embedding, Embeddings};
#[derive(Debug)]
pub struct Embedder {
headers: reqwest::header::HeaderMap,
tokenizer: tiktoken_rs::CoreBPE,
options: EmbedderOptions,
}
use super::rest::{Embedder as RestEmbedder, EmbedderOptions as RestEmbedderOptions};
use super::{DistributionShift, Embeddings};
use crate::vector::error::EmbedErrorKind;
#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
pub struct EmbedderOptions {
pub api_key: Option<String>,
pub embedding_model: EmbeddingModel,
pub dimensions: Option<usize>,
pub distribution: Option<DistributionShift>,
}
impl EmbedderOptions {
pub fn dimensions(&self) -> usize {
if self.embedding_model.supports_overriding_dimensions() {
self.dimensions.unwrap_or(self.embedding_model.default_dimensions())
} else {
self.embedding_model.default_dimensions()
}
}
pub fn query(&self) -> serde_json::Value {
let model = self.embedding_model.name();
let mut query = serde_json::json!({
"model": model,
});
if self.embedding_model.supports_overriding_dimensions() {
if let Some(dimensions) = self.dimensions {
query["dimensions"] = dimensions.into();
}
}
query
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.distribution.or(self.embedding_model.distribution())
}
}
#[derive(
@ -92,15 +116,18 @@ impl EmbeddingModel {
fn distribution(&self) -> Option<DistributionShift> {
match self {
EmbeddingModel::TextEmbeddingAda002 => {
Some(DistributionShift { current_mean: 0.90, current_sigma: 0.08 })
}
EmbeddingModel::TextEmbedding3Large => {
Some(DistributionShift { current_mean: 0.70, current_sigma: 0.1 })
}
EmbeddingModel::TextEmbedding3Small => {
Some(DistributionShift { current_mean: 0.75, current_sigma: 0.1 })
}
EmbeddingModel::TextEmbeddingAda002 => Some(DistributionShift {
current_mean: OrderedFloat(0.90),
current_sigma: OrderedFloat(0.08),
}),
EmbeddingModel::TextEmbedding3Large => Some(DistributionShift {
current_mean: OrderedFloat(0.70),
current_sigma: OrderedFloat(0.1),
}),
EmbeddingModel::TextEmbedding3Small => Some(DistributionShift {
current_mean: OrderedFloat(0.75),
current_sigma: OrderedFloat(0.1),
}),
}
}
@ -117,410 +144,123 @@ pub const OPENAI_EMBEDDINGS_URL: &str = "https://api.openai.com/v1/embeddings";
impl EmbedderOptions {
pub fn with_default_model(api_key: Option<String>) -> Self {
Self { api_key, embedding_model: Default::default(), dimensions: None }
Self { api_key, embedding_model: Default::default(), dimensions: None, distribution: None }
}
pub fn with_embedding_model(api_key: Option<String>, embedding_model: EmbeddingModel) -> Self {
Self { api_key, embedding_model, dimensions: None }
Self { api_key, embedding_model, dimensions: None, distribution: None }
}
}
impl Embedder {
pub fn new_client(&self) -> Result<reqwest::Client, EmbedError> {
reqwest::ClientBuilder::new()
.default_headers(self.headers.clone())
.build()
.map_err(EmbedError::openai_initialize_web_client)
}
pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
let mut headers = reqwest::header::HeaderMap::new();
let mut inferred_api_key = Default::default();
let api_key = options.api_key.as_ref().unwrap_or_else(|| {
inferred_api_key = infer_api_key();
&inferred_api_key
});
headers.insert(
reqwest::header::AUTHORIZATION,
reqwest::header::HeaderValue::from_str(&format!("Bearer {}", api_key))
.map_err(NewEmbedderError::openai_invalid_api_key_format)?,
);
headers.insert(
reqwest::header::CONTENT_TYPE,
reqwest::header::HeaderValue::from_static("application/json"),
);
// looking at the code it is very unclear that this can actually fail.
let tokenizer = tiktoken_rs::cl100k_base().unwrap();
Ok(Self { options, headers, tokenizer })
}
pub async fn embed(
&self,
texts: Vec<String>,
client: &reqwest::Client,
) -> Result<Vec<Embeddings<f32>>, EmbedError> {
let mut tokenized = false;
for attempt in 0..7 {
let result = if tokenized {
self.try_embed_tokenized(&texts, client).await
} else {
self.try_embed(&texts, client).await
};
let retry_duration = match result {
Ok(embeddings) => return Ok(embeddings),
Err(retry) => {
tracing::warn!("Failed: {}", retry.error);
tokenized |= retry.must_tokenize();
retry.into_duration(attempt)
}
}?;
let retry_duration = retry_duration.min(std::time::Duration::from_secs(60)); // don't wait more than a minute
tracing::warn!(
"Attempt #{}, retrying after {}ms.",
attempt,
retry_duration.as_millis()
);
tokio::time::sleep(retry_duration).await;
}
let result = if tokenized {
self.try_embed_tokenized(&texts, client).await
} else {
self.try_embed(&texts, client).await
};
result.map_err(Retry::into_error)
}
async fn check_response(response: reqwest::Response) -> Result<reqwest::Response, Retry> {
if !response.status().is_success() {
match response.status() {
StatusCode::UNAUTHORIZED => {
let error_response: OpenAiErrorResponse = response
.json()
.await
.map_err(EmbedError::openai_unexpected)
.map_err(Retry::retry_later)?;
return Err(Retry::give_up(EmbedError::openai_auth_error(
error_response.error,
)));
}
StatusCode::TOO_MANY_REQUESTS => {
let error_response: OpenAiErrorResponse = response
.json()
.await
.map_err(EmbedError::openai_unexpected)
.map_err(Retry::retry_later)?;
return Err(Retry::rate_limited(EmbedError::openai_too_many_requests(
error_response.error,
)));
}
StatusCode::INTERNAL_SERVER_ERROR
| StatusCode::BAD_GATEWAY
| StatusCode::SERVICE_UNAVAILABLE => {
let error_response: Result<OpenAiErrorResponse, _> = response.json().await;
return Err(Retry::retry_later(EmbedError::openai_internal_server_error(
error_response.ok().map(|error_response| error_response.error),
)));
}
StatusCode::BAD_REQUEST => {
// Most probably, one text contained too many tokens
let error_response: OpenAiErrorResponse = response
.json()
.await
.map_err(EmbedError::openai_unexpected)
.map_err(Retry::retry_later)?;
tracing::warn!("OpenAI: received `BAD_REQUEST`. Input was maybe too long, retrying on tokenized version. For best performance, limit the size of your prompt.");
return Err(Retry::retry_tokenized(EmbedError::openai_too_many_tokens(
error_response.error,
)));
}
code => {
return Err(Retry::retry_later(EmbedError::openai_unhandled_status_code(
code.as_u16(),
)));
}
}
}
Ok(response)
}
async fn try_embed<S: AsRef<str> + serde::Serialize>(
&self,
texts: &[S],
client: &reqwest::Client,
) -> Result<Vec<Embeddings<f32>>, Retry> {
for text in texts {
tracing::trace!("Received prompt: {}", text.as_ref())
}
let request = OpenAiRequest {
model: self.options.embedding_model.name(),
input: texts,
dimensions: self.overriden_dimensions(),
};
let response = client
.post(OPENAI_EMBEDDINGS_URL)
.json(&request)
.send()
.await
.map_err(EmbedError::openai_network)
.map_err(Retry::retry_later)?;
let response = Self::check_response(response).await?;
let response: OpenAiResponse = response
.json()
.await
.map_err(EmbedError::openai_unexpected)
.map_err(Retry::retry_later)?;
tracing::trace!("response: {:?}", response.data);
Ok(response
.data
.into_iter()
.map(|data| Embeddings::from_single_embedding(data.embedding))
.collect())
}
async fn try_embed_tokenized(
&self,
text: &[String],
client: &reqwest::Client,
) -> Result<Vec<Embeddings<f32>>, Retry> {
pub const OVERLAP_SIZE: usize = 200;
let mut all_embeddings = Vec::with_capacity(text.len());
for text in text {
let max_token_count = self.options.embedding_model.max_token();
let encoded = self.tokenizer.encode_ordinary(text.as_str());
let len = encoded.len();
if len < max_token_count {
all_embeddings.append(&mut self.try_embed(&[text], client).await?);
continue;
}
let mut tokens = encoded.as_slice();
let mut embeddings_for_prompt = Embeddings::new(self.dimensions());
while tokens.len() > max_token_count {
let window = &tokens[..max_token_count];
embeddings_for_prompt.push(self.embed_tokens(window, client).await?).unwrap();
tokens = &tokens[max_token_count - OVERLAP_SIZE..];
}
// end of text
embeddings_for_prompt.push(self.embed_tokens(tokens, client).await?).unwrap();
all_embeddings.push(embeddings_for_prompt);
}
Ok(all_embeddings)
}
async fn embed_tokens(
&self,
tokens: &[usize],
client: &reqwest::Client,
) -> Result<Embedding, Retry> {
for attempt in 0..9 {
let duration = match self.try_embed_tokens(tokens, client).await {
Ok(embedding) => return Ok(embedding),
Err(retry) => retry.into_duration(attempt),
}
.map_err(Retry::retry_later)?;
tokio::time::sleep(duration).await;
}
self.try_embed_tokens(tokens, client)
.await
.map_err(|retry| Retry::give_up(retry.into_error()))
}
async fn try_embed_tokens(
&self,
tokens: &[usize],
client: &reqwest::Client,
) -> Result<Embedding, Retry> {
let request = OpenAiTokensRequest {
model: self.options.embedding_model.name(),
input: tokens,
dimensions: self.overriden_dimensions(),
};
let response = client
.post(OPENAI_EMBEDDINGS_URL)
.json(&request)
.send()
.await
.map_err(EmbedError::openai_network)
.map_err(Retry::retry_later)?;
let response = Self::check_response(response).await?;
let mut response: OpenAiResponse = response
.json()
.await
.map_err(EmbedError::openai_unexpected)
.map_err(Retry::retry_later)?;
Ok(response.data.pop().map(|data| data.embedding).unwrap_or_default())
}
pub fn embed_chunks(
&self,
text_chunks: Vec<Vec<String>>,
) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
let rt = tokio::runtime::Builder::new_current_thread()
.enable_io()
.enable_time()
.build()
.map_err(EmbedError::openai_runtime_init)?;
let client = self.new_client()?;
rt.block_on(futures::future::try_join_all(
text_chunks.into_iter().map(|prompts| self.embed(prompts, &client)),
))
}
pub fn chunk_count_hint(&self) -> usize {
10
}
pub fn prompt_count_in_chunk_hint(&self) -> usize {
10
}
pub fn dimensions(&self) -> usize {
if self.options.embedding_model.supports_overriding_dimensions() {
self.options.dimensions.unwrap_or(self.options.embedding_model.default_dimensions())
} else {
self.options.embedding_model.default_dimensions()
}
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.options.embedding_model.distribution()
}
fn overriden_dimensions(&self) -> Option<usize> {
if self.options.embedding_model.supports_overriding_dimensions() {
self.options.dimensions
} else {
None
}
}
}
// retrying in case of failure
struct Retry {
error: EmbedError,
strategy: RetryStrategy,
}
enum RetryStrategy {
GiveUp,
Retry,
RetryTokenized,
RetryAfterRateLimit,
}
impl Retry {
fn give_up(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::GiveUp }
}
fn retry_later(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::Retry }
}
fn retry_tokenized(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::RetryTokenized }
}
fn rate_limited(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::RetryAfterRateLimit }
}
fn into_duration(self, attempt: u32) -> Result<tokio::time::Duration, EmbedError> {
match self.strategy {
RetryStrategy::GiveUp => Err(self.error),
RetryStrategy::Retry => Ok(tokio::time::Duration::from_millis((10u64).pow(attempt))),
RetryStrategy::RetryTokenized => Ok(tokio::time::Duration::from_millis(1)),
RetryStrategy::RetryAfterRateLimit => {
Ok(tokio::time::Duration::from_millis(100 + 10u64.pow(attempt)))
}
}
}
fn must_tokenize(&self) -> bool {
matches!(self.strategy, RetryStrategy::RetryTokenized)
}
fn into_error(self) -> EmbedError {
self.error
}
}
// openai api structs
#[derive(Debug, Serialize)]
struct OpenAiRequest<'a, S: AsRef<str> + serde::Serialize> {
model: &'a str,
input: &'a [S],
#[serde(skip_serializing_if = "Option::is_none")]
dimensions: Option<usize>,
}
#[derive(Debug, Serialize)]
struct OpenAiTokensRequest<'a> {
model: &'a str,
input: &'a [usize],
#[serde(skip_serializing_if = "Option::is_none")]
dimensions: Option<usize>,
}
#[derive(Debug, Deserialize)]
struct OpenAiResponse {
data: Vec<OpenAiEmbedding>,
}
#[derive(Debug, Deserialize)]
struct OpenAiErrorResponse {
error: OpenAiError,
}
#[derive(Debug, Deserialize)]
pub struct OpenAiError {
message: String,
// type: String,
code: Option<String>,
}
impl Display for OpenAiError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match &self.code {
Some(code) => write!(f, "{} ({})", self.message, code),
None => write!(f, "{}", self.message),
}
}
}
#[derive(Debug, Deserialize)]
struct OpenAiEmbedding {
embedding: Embedding,
// object: String,
// index: usize,
}
fn infer_api_key() -> String {
std::env::var("MEILI_OPENAI_API_KEY")
.or_else(|_| std::env::var("OPENAI_API_KEY"))
.unwrap_or_default()
}
#[derive(Debug)]
pub struct Embedder {
tokenizer: tiktoken_rs::CoreBPE,
rest_embedder: RestEmbedder,
options: EmbedderOptions,
}
impl Embedder {
pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
let mut inferred_api_key = Default::default();
let api_key = options.api_key.as_ref().unwrap_or_else(|| {
inferred_api_key = infer_api_key();
&inferred_api_key
});
let rest_embedder = RestEmbedder::new(RestEmbedderOptions {
api_key: Some(api_key.clone()),
distribution: None,
dimensions: Some(options.dimensions()),
url: OPENAI_EMBEDDINGS_URL.to_owned(),
query: options.query(),
input_field: vec!["input".to_owned()],
input_type: crate::vector::rest::InputType::TextArray,
path_to_embeddings: vec!["data".to_owned()],
embedding_object: vec!["embedding".to_owned()],
})?;
// looking at the code it is very unclear that this can actually fail.
let tokenizer = tiktoken_rs::cl100k_base().unwrap();
Ok(Self { options, rest_embedder, tokenizer })
}
pub fn embed(&self, texts: Vec<String>) -> Result<Vec<Embeddings<f32>>, EmbedError> {
match self.rest_embedder.embed_ref(&texts) {
Ok(embeddings) => Ok(embeddings),
Err(EmbedError { kind: EmbedErrorKind::RestBadRequest(error), fault: _ }) => {
tracing::warn!(error=?error, "OpenAI: received `BAD_REQUEST`. Input was maybe too long, retrying on tokenized version. For best performance, limit the size of your document template.");
self.try_embed_tokenized(&texts)
}
Err(error) => Err(error),
}
}
fn try_embed_tokenized(&self, text: &[String]) -> Result<Vec<Embeddings<f32>>, EmbedError> {
pub const OVERLAP_SIZE: usize = 200;
let mut all_embeddings = Vec::with_capacity(text.len());
for text in text {
let max_token_count = self.options.embedding_model.max_token();
let encoded = self.tokenizer.encode_ordinary(text.as_str());
let len = encoded.len();
if len < max_token_count {
all_embeddings.append(&mut self.rest_embedder.embed_ref(&[text])?);
continue;
}
let mut tokens = encoded.as_slice();
let mut embeddings_for_prompt = Embeddings::new(self.dimensions());
while tokens.len() > max_token_count {
let window = &tokens[..max_token_count];
let embedding = self.rest_embedder.embed_tokens(window)?;
embeddings_for_prompt.append(embedding.into_inner()).map_err(|got| {
EmbedError::openai_unexpected_dimension(self.dimensions(), got.len())
})?;
tokens = &tokens[max_token_count - OVERLAP_SIZE..];
}
// end of text
let embedding = self.rest_embedder.embed_tokens(tokens)?;
embeddings_for_prompt.append(embedding.into_inner()).map_err(|got| {
EmbedError::openai_unexpected_dimension(self.dimensions(), got.len())
})?;
all_embeddings.push(embeddings_for_prompt);
}
Ok(all_embeddings)
}
pub fn embed_chunks(
&self,
text_chunks: Vec<Vec<String>>,
threads: &rayon::ThreadPool,
) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
threads.install(move || {
text_chunks.into_par_iter().map(move |chunk| self.embed(chunk)).collect()
})
}
pub fn chunk_count_hint(&self) -> usize {
self.rest_embedder.chunk_count_hint()
}
pub fn prompt_count_in_chunk_hint(&self) -> usize {
self.rest_embedder.prompt_count_in_chunk_hint()
}
pub fn dimensions(&self) -> usize {
self.options.dimensions()
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.options.distribution()
}
}

373
milli/src/vector/rest.rs Normal file
View File

@ -0,0 +1,373 @@
use deserr::Deserr;
use rayon::iter::{IntoParallelIterator as _, ParallelIterator as _};
use serde::{Deserialize, Serialize};
use super::{
DistributionShift, EmbedError, Embedding, Embeddings, NewEmbedderError, REQUEST_PARALLELISM,
};
// retrying in case of failure
pub struct Retry {
pub error: EmbedError,
strategy: RetryStrategy,
}
pub enum RetryStrategy {
GiveUp,
Retry,
RetryTokenized,
RetryAfterRateLimit,
}
impl Retry {
pub fn give_up(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::GiveUp }
}
pub fn retry_later(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::Retry }
}
pub fn retry_tokenized(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::RetryTokenized }
}
pub fn rate_limited(error: EmbedError) -> Self {
Self { error, strategy: RetryStrategy::RetryAfterRateLimit }
}
pub fn into_duration(self, attempt: u32) -> Result<std::time::Duration, EmbedError> {
match self.strategy {
RetryStrategy::GiveUp => Err(self.error),
RetryStrategy::Retry => Ok(std::time::Duration::from_millis((10u64).pow(attempt))),
RetryStrategy::RetryTokenized => Ok(std::time::Duration::from_millis(1)),
RetryStrategy::RetryAfterRateLimit => {
Ok(std::time::Duration::from_millis(100 + 10u64.pow(attempt)))
}
}
}
pub fn must_tokenize(&self) -> bool {
matches!(self.strategy, RetryStrategy::RetryTokenized)
}
pub fn into_error(self) -> EmbedError {
self.error
}
}
#[derive(Debug)]
pub struct Embedder {
client: ureq::Agent,
options: EmbedderOptions,
bearer: Option<String>,
dimensions: usize,
}
#[derive(Debug, Clone, PartialEq, Eq, Deserialize, Serialize)]
pub struct EmbedderOptions {
pub api_key: Option<String>,
pub distribution: Option<DistributionShift>,
pub dimensions: Option<usize>,
pub url: String,
pub query: serde_json::Value,
pub input_field: Vec<String>,
// path to the array of embeddings
pub path_to_embeddings: Vec<String>,
// shape of a single embedding
pub embedding_object: Vec<String>,
pub input_type: InputType,
}
impl Default for EmbedderOptions {
fn default() -> Self {
Self {
url: Default::default(),
query: Default::default(),
input_field: vec!["input".into()],
path_to_embeddings: vec!["data".into()],
embedding_object: vec!["embedding".into()],
input_type: InputType::Text,
api_key: None,
distribution: None,
dimensions: None,
}
}
}
impl std::hash::Hash for EmbedderOptions {
fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
self.api_key.hash(state);
self.distribution.hash(state);
self.dimensions.hash(state);
self.url.hash(state);
// skip hashing the query
// collisions in regular usage should be minimal,
// and the list is limited to 256 values anyway
self.input_field.hash(state);
self.path_to_embeddings.hash(state);
self.embedding_object.hash(state);
self.input_type.hash(state);
}
}
#[derive(Debug, Clone, Copy, Deserialize, Serialize, PartialEq, Eq, Hash, Deserr)]
#[serde(rename_all = "camelCase")]
#[deserr(rename_all = camelCase, deny_unknown_fields)]
pub enum InputType {
Text,
TextArray,
}
impl Embedder {
pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
let bearer = options.api_key.as_deref().map(|api_key| format!("Bearer {api_key}"));
let client = ureq::AgentBuilder::new()
.max_idle_connections(REQUEST_PARALLELISM * 2)
.max_idle_connections_per_host(REQUEST_PARALLELISM * 2)
.build();
let dimensions = if let Some(dimensions) = options.dimensions {
dimensions
} else {
infer_dimensions(&client, &options, bearer.as_deref())?
};
Ok(Self { client, dimensions, options, bearer })
}
pub fn embed(&self, texts: Vec<String>) -> Result<Vec<Embeddings<f32>>, EmbedError> {
embed(&self.client, &self.options, self.bearer.as_deref(), texts.as_slice(), texts.len())
}
pub fn embed_ref<S>(&self, texts: &[S]) -> Result<Vec<Embeddings<f32>>, EmbedError>
where
S: AsRef<str> + Serialize,
{
embed(&self.client, &self.options, self.bearer.as_deref(), texts, texts.len())
}
pub fn embed_tokens(&self, tokens: &[usize]) -> Result<Embeddings<f32>, EmbedError> {
let mut embeddings = embed(&self.client, &self.options, self.bearer.as_deref(), tokens, 1)?;
// unwrap: guaranteed that embeddings.len() == 1, otherwise the previous line terminated in error
Ok(embeddings.pop().unwrap())
}
pub fn embed_chunks(
&self,
text_chunks: Vec<Vec<String>>,
threads: &rayon::ThreadPool,
) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
threads.install(move || {
text_chunks.into_par_iter().map(move |chunk| self.embed(chunk)).collect()
})
}
pub fn chunk_count_hint(&self) -> usize {
super::REQUEST_PARALLELISM
}
pub fn prompt_count_in_chunk_hint(&self) -> usize {
match self.options.input_type {
InputType::Text => 1,
InputType::TextArray => 10,
}
}
pub fn dimensions(&self) -> usize {
self.dimensions
}
pub fn distribution(&self) -> Option<DistributionShift> {
self.options.distribution
}
}
fn infer_dimensions(
client: &ureq::Agent,
options: &EmbedderOptions,
bearer: Option<&str>,
) -> Result<usize, NewEmbedderError> {
let v = embed(client, options, bearer, ["test"].as_slice(), 1)
.map_err(NewEmbedderError::could_not_determine_dimension)?;
// unwrap: guaranteed that v.len() == 1, otherwise the previous line terminated in error
Ok(v.first().unwrap().dimension())
}
fn embed<S>(
client: &ureq::Agent,
options: &EmbedderOptions,
bearer: Option<&str>,
inputs: &[S],
expected_count: usize,
) -> Result<Vec<Embeddings<f32>>, EmbedError>
where
S: Serialize,
{
let request = client.post(&options.url);
let request =
if let Some(bearer) = bearer { request.set("Authorization", bearer) } else { request };
let request = request.set("Content-Type", "application/json");
let input_value = match options.input_type {
InputType::Text => serde_json::json!(inputs.first()),
InputType::TextArray => serde_json::json!(inputs),
};
let body = match options.input_field.as_slice() {
[] => {
// inject input in body
input_value
}
[input] => {
let mut body = options.query.clone();
body.as_object_mut()
.ok_or_else(|| {
EmbedError::rest_not_an_object(
options.query.clone(),
options.input_field.clone(),
)
})?
.insert(input.clone(), input_value);
body
}
[path @ .., input] => {
let mut body = options.query.clone();
let mut current_value = &mut body;
for component in path {
current_value = current_value
.as_object_mut()
.ok_or_else(|| {
EmbedError::rest_not_an_object(
options.query.clone(),
options.input_field.clone(),
)
})?
.entry(component.clone())
.or_insert(serde_json::json!({}));
}
current_value.as_object_mut().unwrap().insert(input.clone(), input_value);
body
}
};
for attempt in 0..7 {
let response = request.clone().send_json(&body);
let result = check_response(response);
let retry_duration = match result {
Ok(response) => return response_to_embedding(response, options, expected_count),
Err(retry) => {
tracing::warn!("Failed: {}", retry.error);
retry.into_duration(attempt)
}
}?;
let retry_duration = retry_duration.min(std::time::Duration::from_secs(60)); // don't wait more than a minute
tracing::warn!("Attempt #{}, retrying after {}ms.", attempt, retry_duration.as_millis());
std::thread::sleep(retry_duration);
}
let response = request.send_json(&body);
let result = check_response(response);
result
.map_err(Retry::into_error)
.and_then(|response| response_to_embedding(response, options, expected_count))
}
fn check_response(response: Result<ureq::Response, ureq::Error>) -> Result<ureq::Response, Retry> {
match response {
Ok(response) => Ok(response),
Err(ureq::Error::Status(code, response)) => {
let error_response: Option<String> = response.into_string().ok();
Err(match code {
401 => Retry::give_up(EmbedError::rest_unauthorized(error_response)),
429 => Retry::rate_limited(EmbedError::rest_too_many_requests(error_response)),
400 => Retry::give_up(EmbedError::rest_bad_request(error_response)),
500..=599 => {
Retry::retry_later(EmbedError::rest_internal_server_error(code, error_response))
}
402..=499 => {
Retry::give_up(EmbedError::rest_other_status_code(code, error_response))
}
_ => Retry::retry_later(EmbedError::rest_other_status_code(code, error_response)),
})
}
Err(ureq::Error::Transport(transport)) => {
Err(Retry::retry_later(EmbedError::rest_network(transport)))
}
}
}
fn response_to_embedding(
response: ureq::Response,
options: &EmbedderOptions,
expected_count: usize,
) -> Result<Vec<Embeddings<f32>>, EmbedError> {
let response: serde_json::Value =
response.into_json().map_err(EmbedError::rest_response_deserialization)?;
let mut current_value = &response;
for component in &options.path_to_embeddings {
let component = component.as_ref();
current_value = current_value.get(component).ok_or_else(|| {
EmbedError::rest_response_missing_embeddings(
response.clone(),
component,
&options.path_to_embeddings,
)
})?;
}
let embeddings = match options.input_type {
InputType::Text => {
for component in &options.embedding_object {
current_value = current_value.get(component).ok_or_else(|| {
EmbedError::rest_response_missing_embeddings(
response.clone(),
component,
&options.embedding_object,
)
})?;
}
let embeddings = current_value.to_owned();
let embeddings: Embedding =
serde_json::from_value(embeddings).map_err(EmbedError::rest_response_format)?;
vec![Embeddings::from_single_embedding(embeddings)]
}
InputType::TextArray => {
let empty = vec![];
let values = current_value.as_array().unwrap_or(&empty);
let mut embeddings: Vec<Embeddings<f32>> = Vec::with_capacity(expected_count);
for value in values {
let mut current_value = value;
for component in &options.embedding_object {
current_value = current_value.get(component).ok_or_else(|| {
EmbedError::rest_response_missing_embeddings(
response.clone(),
component,
&options.embedding_object,
)
})?;
}
let embedding = current_value.to_owned();
let embedding: Embedding =
serde_json::from_value(embedding).map_err(EmbedError::rest_response_format)?;
embeddings.push(Embeddings::from_single_embedding(embedding));
}
embeddings
}
};
if embeddings.len() != expected_count {
return Err(EmbedError::rest_response_embedding_count(expected_count, embeddings.len()));
}
Ok(embeddings)
}

313
milli/src/vector/settings.rs
View File

@ -1,7 +1,8 @@
use deserr::Deserr;
use serde::{Deserialize, Serialize};
use super::openai;
use super::rest::InputType;
use super::{ollama, openai, DistributionShift};
use crate::prompt::PromptData;
use crate::update::Setting;
use crate::vector::EmbeddingConfig;
@ -29,6 +30,27 @@ pub struct EmbeddingSettings {
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub document_template: Setting<String>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub url: Setting<String>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub query: Setting<serde_json::Value>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub input_field: Setting<Vec<String>>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub path_to_embeddings: Setting<Vec<String>>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub embedding_object: Setting<Vec<String>>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub input_type: Setting<InputType>,
#[serde(default, skip_serializing_if = "Setting::is_not_set")]
#[deserr(default)]
pub distribution: Setting<DistributionShift>,
}
pub fn check_unset<T>(
@ -75,16 +97,53 @@ impl EmbeddingSettings {
pub const DIMENSIONS: &'static str = "dimensions";
pub const DOCUMENT_TEMPLATE: &'static str = "documentTemplate";
pub const URL: &'static str = "url";
pub const QUERY: &'static str = "query";
pub const INPUT_FIELD: &'static str = "inputField";
pub const PATH_TO_EMBEDDINGS: &'static str = "pathToEmbeddings";
pub const EMBEDDING_OBJECT: &'static str = "embeddingObject";
pub const INPUT_TYPE: &'static str = "inputType";
pub const DISTRIBUTION: &'static str = "distribution";
pub fn allowed_sources_for_field(field: &'static str) -> &'static [EmbedderSource] {
match field {
Self::SOURCE => {
&[EmbedderSource::HuggingFace, EmbedderSource::OpenAi, EmbedderSource::UserProvided]
Self::SOURCE => &[
EmbedderSource::HuggingFace,
EmbedderSource::OpenAi,
EmbedderSource::UserProvided,
EmbedderSource::Rest,
EmbedderSource::Ollama,
],
Self::MODEL => {
&[EmbedderSource::HuggingFace, EmbedderSource::OpenAi, EmbedderSource::Ollama]
}
Self::MODEL => &[EmbedderSource::HuggingFace, EmbedderSource::OpenAi],
Self::REVISION => &[EmbedderSource::HuggingFace],
Self::API_KEY => &[EmbedderSource::OpenAi],
Self::DIMENSIONS => &[EmbedderSource::OpenAi, EmbedderSource::UserProvided],
Self::DOCUMENT_TEMPLATE => &[EmbedderSource::HuggingFace, EmbedderSource::OpenAi],
Self::API_KEY => {
&[EmbedderSource::OpenAi, EmbedderSource::Ollama, EmbedderSource::Rest]
}
Self::DIMENSIONS => {
&[EmbedderSource::OpenAi, EmbedderSource::UserProvided, EmbedderSource::Rest]
}
Self::DOCUMENT_TEMPLATE => &[
EmbedderSource::HuggingFace,
EmbedderSource::OpenAi,
EmbedderSource::Ollama,
EmbedderSource::Rest,
],
Self::URL => &[EmbedderSource::Ollama, EmbedderSource::Rest],
Self::QUERY => &[EmbedderSource::Rest],
Self::INPUT_FIELD => &[EmbedderSource::Rest],
Self::PATH_TO_EMBEDDINGS => &[EmbedderSource::Rest],
Self::EMBEDDING_OBJECT => &[EmbedderSource::Rest],
Self::INPUT_TYPE => &[EmbedderSource::Rest],
Self::DISTRIBUTION => &[
EmbedderSource::HuggingFace,
EmbedderSource::Ollama,
EmbedderSource::OpenAi,
EmbedderSource::Rest,
EmbedderSource::UserProvided,
],
_other => unreachable!("unknown field"),
}
}
@ -97,11 +156,37 @@ impl EmbeddingSettings {
Self::API_KEY,
Self::DOCUMENT_TEMPLATE,
Self::DIMENSIONS,
Self::DISTRIBUTION,
],
EmbedderSource::HuggingFace => &[
Self::SOURCE,
Self::MODEL,
Self::REVISION,
Self::DOCUMENT_TEMPLATE,
Self::DISTRIBUTION,
],
EmbedderSource::Ollama => &[
Self::SOURCE,
Self::MODEL,
Self::DOCUMENT_TEMPLATE,
Self::URL,
Self::API_KEY,
Self::DISTRIBUTION,
],
EmbedderSource::UserProvided => &[Self::SOURCE, Self::DIMENSIONS, Self::DISTRIBUTION],
EmbedderSource::Rest => &[
Self::SOURCE,
Self::API_KEY,
Self::DIMENSIONS,
Self::DOCUMENT_TEMPLATE,
Self::URL,
Self::QUERY,
Self::INPUT_FIELD,
Self::PATH_TO_EMBEDDINGS,
Self::EMBEDDING_OBJECT,
Self::INPUT_TYPE,
Self::DISTRIBUTION,
],
EmbedderSource::HuggingFace => {
&[Self::SOURCE, Self::MODEL, Self::REVISION, Self::DOCUMENT_TEMPLATE]
}
EmbedderSource::UserProvided => &[Self::SOURCE, Self::DIMENSIONS],
}
}
@ -125,6 +210,66 @@ impl EmbeddingSettings {
*model = Setting::Set(openai::EmbeddingModel::default().name().to_owned())
}
}
pub(crate) fn apply_and_need_reindex(
old: &mut Setting<EmbeddingSettings>,
new: Setting<EmbeddingSettings>,
) -> bool {
match (old, new) {
(
Setting::Set(EmbeddingSettings {
source: old_source,
model: old_model,
revision: old_revision,
api_key: old_api_key,
dimensions: old_dimensions,
document_template: old_document_template,
url: old_url,
query: old_query,
input_field: old_input_field,
path_to_embeddings: old_path_to_embeddings,
embedding_object: old_embedding_object,
input_type: old_input_type,
distribution: old_distribution,
}),
Setting::Set(EmbeddingSettings {
source: new_source,
model: new_model,
revision: new_revision,
api_key: new_api_key,
dimensions: new_dimensions,
document_template: new_document_template,
url: new_url,
query: new_query,
input_field: new_input_field,
path_to_embeddings: new_path_to_embeddings,
embedding_object: new_embedding_object,
input_type: new_input_type,
distribution: new_distribution,
}),
) => {
let mut needs_reindex = false;
needs_reindex |= old_source.apply(new_source);
needs_reindex |= old_model.apply(new_model);
needs_reindex |= old_revision.apply(new_revision);
needs_reindex |= old_dimensions.apply(new_dimensions);
needs_reindex |= old_document_template.apply(new_document_template);
needs_reindex |= old_url.apply(new_url);
needs_reindex |= old_query.apply(new_query);
needs_reindex |= old_input_field.apply(new_input_field);
needs_reindex |= old_path_to_embeddings.apply(new_path_to_embeddings);
needs_reindex |= old_embedding_object.apply(new_embedding_object);
needs_reindex |= old_input_type.apply(new_input_type);
old_distribution.apply(new_distribution);
old_api_key.apply(new_api_key);
needs_reindex
}
(Setting::Reset, Setting::Reset) | (_, Setting::NotSet) => false,
_ => true,
}
}
}
#[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq, Deserr)]
@ -134,7 +279,9 @@ pub enum EmbedderSource {
#[default]
OpenAi,
HuggingFace,
Ollama,
UserProvided,
Rest,
}
impl std::fmt::Display for EmbedderSource {
@ -143,38 +290,13 @@ impl std::fmt::Display for EmbedderSource {
EmbedderSource::OpenAi => "openAi",
EmbedderSource::HuggingFace => "huggingFace",
EmbedderSource::UserProvided => "userProvided",
EmbedderSource::Ollama => "ollama",
EmbedderSource::Rest => "rest",
};
f.write_str(s)
}
}
impl EmbeddingSettings {
pub fn apply(&mut self, new: Self) {
let EmbeddingSettings { source, model, revision, api_key, dimensions, document_template } =
new;
let old_source = self.source;
self.source.apply(source);
// Reinitialize the whole setting object on a source change
if old_source != self.source {
*self = EmbeddingSettings {
source,
model,
revision,
api_key,
dimensions,
document_template,
};
return;
}
self.model.apply(model);
self.revision.apply(revision);
self.api_key.apply(api_key);
self.dimensions.apply(dimensions);
self.document_template.apply(document_template);
}
}
impl From<EmbeddingConfig> for EmbeddingSettings {
fn from(value: EmbeddingConfig) -> Self {
let EmbeddingConfig { embedder_options, prompt } = value;
@ -186,6 +308,13 @@ impl From<EmbeddingConfig> for EmbeddingSettings {
api_key: Setting::NotSet,
dimensions: Setting::NotSet,
document_template: Setting::Set(prompt.template),
url: Setting::NotSet,
query: Setting::NotSet,
input_field: Setting::NotSet,
path_to_embeddings: Setting::NotSet,
embedding_object: Setting::NotSet,
input_type: Setting::NotSet,
distribution: options.distribution.map(Setting::Set).unwrap_or_default(),
},
super::EmbedderOptions::OpenAi(options) => Self {
source: Setting::Set(EmbedderSource::OpenAi),
@ -194,6 +323,28 @@ impl From<EmbeddingConfig> for EmbeddingSettings {
api_key: options.api_key.map(Setting::Set).unwrap_or_default(),
dimensions: options.dimensions.map(Setting::Set).unwrap_or_default(),
document_template: Setting::Set(prompt.template),
url: Setting::NotSet,
query: Setting::NotSet,
input_field: Setting::NotSet,
path_to_embeddings: Setting::NotSet,
embedding_object: Setting::NotSet,
input_type: Setting::NotSet,
distribution: options.distribution.map(Setting::Set).unwrap_or_default(),
},
super::EmbedderOptions::Ollama(options) => Self {
source: Setting::Set(EmbedderSource::Ollama),
model: Setting::Set(options.embedding_model.to_owned()),
revision: Setting::NotSet,
api_key: Setting::NotSet,
dimensions: Setting::NotSet,
document_template: Setting::Set(prompt.template),
url: Setting::NotSet,
query: Setting::NotSet,
input_field: Setting::NotSet,
path_to_embeddings: Setting::NotSet,
embedding_object: Setting::NotSet,
input_type: Setting::NotSet,
distribution: options.distribution.map(Setting::Set).unwrap_or_default(),
},
super::EmbedderOptions::UserProvided(options) => Self {
source: Setting::Set(EmbedderSource::UserProvided),
@ -202,6 +353,38 @@ impl From<EmbeddingConfig> for EmbeddingSettings {
api_key: Setting::NotSet,
dimensions: Setting::Set(options.dimensions),
document_template: Setting::NotSet,
url: Setting::NotSet,
query: Setting::NotSet,
input_field: Setting::NotSet,
path_to_embeddings: Setting::NotSet,
embedding_object: Setting::NotSet,
input_type: Setting::NotSet,
distribution: options.distribution.map(Setting::Set).unwrap_or_default(),
},
super::EmbedderOptions::Rest(super::rest::EmbedderOptions {
api_key,
dimensions,
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
}) => Self {
source: Setting::Set(EmbedderSource::Rest),
model: Setting::NotSet,
revision: Setting::NotSet,
api_key: api_key.map(Setting::Set).unwrap_or_default(),
dimensions: dimensions.map(Setting::Set).unwrap_or_default(),
document_template: Setting::Set(prompt.template),
url: Setting::Set(url),
query: Setting::Set(query),
input_field: Setting::Set(input_field),
path_to_embeddings: Setting::Set(path_to_embeddings),
embedding_object: Setting::Set(embedding_object),
input_type: Setting::Set(input_type),
distribution: distribution.map(Setting::Set).unwrap_or_default(),
},
}
}
@ -210,8 +393,22 @@ impl From<EmbeddingConfig> for EmbeddingSettings {
impl From<EmbeddingSettings> for EmbeddingConfig {
fn from(value: EmbeddingSettings) -> Self {
let mut this = Self::default();
let EmbeddingSettings { source, model, revision, api_key, dimensions, document_template } =
value;
let EmbeddingSettings {
source,
model,
revision,
api_key,
dimensions,
document_template,
url,
query,
input_field,
path_to_embeddings,
embedding_object,
input_type,
distribution,
} = value;
if let Some(source) = source.set() {
match source {
EmbedderSource::OpenAi => {
@ -227,8 +424,22 @@ impl From<EmbeddingSettings> for EmbeddingConfig {
if let Some(dimensions) = dimensions.set() {
options.dimensions = Some(dimensions);
}
options.distribution = distribution.set();
this.embedder_options = super::EmbedderOptions::OpenAi(options);
}
EmbedderSource::Ollama => {
let mut options: ollama::EmbedderOptions =
super::ollama::EmbedderOptions::with_default_model(
api_key.set(),
url.set(),
);
if let Some(model) = model.set() {
options.embedding_model = model;
}
options.distribution = distribution.set();
this.embedder_options = super::EmbedderOptions::Ollama(options);
}
EmbedderSource::HuggingFace => {
let mut options = super::hf::EmbedderOptions::default();
if let Some(model) = model.set() {
@ -243,14 +454,36 @@ impl From<EmbeddingSettings> for EmbeddingConfig {
if let Some(revision) = revision.set() {
options.revision = Some(revision);
}
options.distribution = distribution.set();
this.embedder_options = super::EmbedderOptions::HuggingFace(options);
}
EmbedderSource::UserProvided => {
this.embedder_options =
super::EmbedderOptions::UserProvided(super::manual::EmbedderOptions {
dimensions: dimensions.set().unwrap(),
distribution: distribution.set(),
});
}
EmbedderSource::Rest => {
let embedder_options = super::rest::EmbedderOptions::default();
this.embedder_options =
super::EmbedderOptions::Rest(super::rest::EmbedderOptions {
api_key: api_key.set(),
dimensions: dimensions.set(),
url: url.set().unwrap(),
query: query.set().unwrap_or(embedder_options.query),
input_field: input_field.set().unwrap_or(embedder_options.input_field),
path_to_embeddings: path_to_embeddings
.set()
.unwrap_or(embedder_options.path_to_embeddings),
embedding_object: embedding_object
.set()
.unwrap_or(embedder_options.embedding_object),
input_type: input_type.set().unwrap_or(embedder_options.input_type),
distribution: distribution.set(),
})
}
}
}

94
workloads/settings-add-remove-filters.json Normal file
View File

@ -0,0 +1,94 @@
{
"name": "settings-add-remove-filters.json",
"run_count": 2,
"extra_cli_args": [
"--max-indexing-threads=4"
],
"assets": {
"150k-people.json": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/150k-people.json",
"sha256": "28c359a0956958af0ba204ec11bad3045a0864a10b4838914fea25a01724f84b"
}
},
"commands": [
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
],
"filterableAttributes": [
"city",
"region",
"country_code"
],
"dictionary": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
],
"stopWords": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
]
}
},
"synchronous": "DontWait"
},
{
"route": "indexes/peoples/documents",
"method": "POST",
"body": {
"asset": "150k-people.json"
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"filterableAttributes": [
"city",
"region",
"country_code",
"featured_job_title",
"featured_job_organization_name"
]
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"filterableAttributes": [
"city",
"region",
"country_code"
]
}
},
"synchronous": "WaitForTask"
}
]
}

86
workloads/settings-proximity-precision.json Normal file
View File

@ -0,0 +1,86 @@
{
"name": "settings-proximity-precision.json",
"run_count": 2,
"extra_cli_args": [
"--max-indexing-threads=4"
],
"assets": {
"150k-people.json": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/150k-people.json",
"sha256": "28c359a0956958af0ba204ec11bad3045a0864a10b4838914fea25a01724f84b"
}
},
"commands": [
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
],
"filterableAttributes": [
"city",
"region",
"country_code",
"featured_job_title",
"featured_job_organization_name"
],
"dictionary": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
],
"stopWords": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
]
}
},
"synchronous": "DontWait"
},
{
"route": "indexes/peoples/documents",
"method": "POST",
"body": {
"asset": "150k-people.json"
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"proximityPrecision": "byAttribute"
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"proximityPrecision": "byWord"
}
},
"synchronous": "WaitForTask"
}
]
}

114
workloads/settings-remove-add-swap-searchable.json Normal file
View File

@ -0,0 +1,114 @@
{
"name": "settings-remove-add-swap-searchable.json",
"run_count": 2,
"extra_cli_args": [
"--max-indexing-threads=4"
],
"assets": {
"150k-people.json": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/150k-people.json",
"sha256": "28c359a0956958af0ba204ec11bad3045a0864a10b4838914fea25a01724f84b"
}
},
"commands": [
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
],
"filterableAttributes": [
"city",
"region",
"country_code",
"featured_job_title",
"featured_job_organization_name"
],
"dictionary": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
],
"stopWords": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
]
}
},
"synchronous": "DontWait"
},
{
"route": "indexes/peoples/documents",
"method": "POST",
"body": {
"asset": "150k-people.json"
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_organization_name"
]
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
]
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"first_name",
"last_name",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
]
}
},
"synchronous": "WaitForTask"
}
]
}

115
workloads/settings-typo.json Normal file
View File

@ -0,0 +1,115 @@
{
"name": "settings-typo.json",
"run_count": 2,
"extra_cli_args": [
"--max-indexing-threads=4"
],
"assets": {
"150k-people.json": {
"local_location": null,
"remote_location": "https://milli-benchmarks.fra1.digitaloceanspaces.com/bench/datasets/150k-people.json",
"sha256": "28c359a0956958af0ba204ec11bad3045a0864a10b4838914fea25a01724f84b"
}
},
"commands": [
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"searchableAttributes": [
"last_name",
"first_name",
"featured_job_title",
"featured_job_organization_name",
"facebook_url",
"twitter_url",
"linkedin_url"
],
"filterableAttributes": [
"city",
"region",
"country_code",
"featured_job_title",
"featured_job_organization_name"
],
"dictionary": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
],
"stopWords": [
"https://",
"http://",
"www.",
"crunchbase.com",
"facebook.com",
"twitter.com",
"linkedin.com"
]
}
},
"synchronous": "DontWait"
},
{
"route": "indexes/peoples/documents",
"method": "POST",
"body": {
"asset": "150k-people.json"
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"typoTolerance": {
"disableOnAttributes": ["featured_job_organization_name"]
}
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"typoTolerance": {
"disableOnAttributes": []
}
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"typoTolerance": {
"disableOnWords": ["Ben","Elowitz","Kevin","Flaherty", "Ron", "Dustin", "Owen", "Chris", "Mark", "Matt", "Peter", "Van", "Head", "of"]
}
}
},
"synchronous": "WaitForTask"
},
{
"route": "indexes/peoples/settings",
"method": "PATCH",
"body": {
"inline": {
"typoTolerance": {
"disableOnWords": []
}
}
},
"synchronous": "WaitForTask"
}
]
}

312
xtask/src/bench/dashboard.rs
View File

@ -11,157 +11,179 @@ use super::client::Client;
use super::env_info;
use super::workload::Workload;
pub async fn cancel_on_ctrl_c(
invocation_uuid: Uuid,
dashboard_client: Client,
abort_handle: AbortHandle,
) {
tracing::info!("press Ctrl-C to cancel the invocation");
match ctrl_c().await {
Ok(()) => {
tracing::info!(%invocation_uuid, "received Ctrl-C, cancelling invocation");
mark_as_failed(dashboard_client, invocation_uuid, None).await;
abort_handle.abort();
#[derive(Debug, Clone)]
pub enum DashboardClient {
Client(Client),
Dry,
}
impl DashboardClient {
pub fn new(dashboard_url: &str, api_key: Option<&str>) -> anyhow::Result<Self> {
let dashboard_client = Client::new(
Some(format!("{}/api/v1", dashboard_url)),
api_key,
Some(std::time::Duration::from_secs(60)),
)?;
Ok(Self::Client(dashboard_client))
}
pub fn new_dry() -> Self {
Self::Dry
}
pub async fn send_machine_info(&self, env: &env_info::Environment) -> anyhow::Result<()> {
let Self::Client(dashboard_client) = self else { return Ok(()) };
let response = dashboard_client
.put("machine")
.json(&json!({"hostname": env.hostname}))
.send()
.await
.context("sending machine information")?;
if !response.status().is_success() {
bail!(
"could not send machine information: {} {}",
response.status(),
response.text().await.unwrap_or_else(|_| "unknown".into())
);
}
Err(error) => tracing::warn!(
error = &error as &dyn std::error::Error,
"failed to listen to Ctrl-C signal, invocation won't be canceled on Ctrl-C"
),
Ok(())
}
}
pub async fn mark_as_failed(
dashboard_client: Client,
invocation_uuid: Uuid,
failure_reason: Option<String>,
) {
let response = dashboard_client
.post("cancel-invocation")
.json(&json!({
"invocation_uuid": invocation_uuid,
"failure_reason": failure_reason,
}))
.send()
.await;
let response = match response {
Ok(response) => response,
Err(response_error) => {
tracing::error!(error = &response_error as &dyn std::error::Error, %invocation_uuid, "could not mark invocation as failed");
return;
pub async fn create_invocation(
&self,
build_info: build_info::BuildInfo,
commit_message: &str,
env: env_info::Environment,
max_workloads: usize,
reason: Option<&str>,
) -> anyhow::Result<Uuid> {
let Self::Client(dashboard_client) = self else { return Ok(Uuid::now_v7()) };
let response = dashboard_client
.put("invocation")
.json(&json!({
"commit": {
"sha1": build_info.commit_sha1,
"message": commit_message,
"commit_date": build_info.commit_timestamp,
"branch": build_info.branch,
"tag": build_info.describe.and_then(|describe| describe.as_tag()),
},
"machine_hostname": env.hostname,
"max_workloads": max_workloads,
"reason": reason
}))
.send()
.await
.context("sending invocation")?;
if !response.status().is_success() {
bail!(
"could not send new invocation: {}",
response.text().await.unwrap_or_else(|_| "unknown".into())
);
}
};
if !response.status().is_success() {
tracing::error!(
%invocation_uuid,
"could not mark invocation as failed: {}",
response.text().await.unwrap()
);
return;
}
tracing::warn!(%invocation_uuid, "marked invocation as failed or canceled");
}
pub async fn send_machine_info(
dashboard_client: &Client,
env: &env_info::Environment,
) -> anyhow::Result<()> {
let response = dashboard_client
.put("machine")
.json(&json!({"hostname": env.hostname}))
.send()
.await
.context("sending machine information")?;
if !response.status().is_success() {
bail!(
"could not send machine information: {} {}",
response.status(),
response.text().await.unwrap_or_else(|_| "unknown".into())
);
}
Ok(())
}
pub async fn create_invocation(
dashboard_client: &Client,
build_info: build_info::BuildInfo,
commit_message: &str,
env: env_info::Environment,
max_workloads: usize,
reason: Option<&str>,
) -> anyhow::Result<Uuid> {
let response = dashboard_client
.put("invocation")
.json(&json!({
"commit": {
"sha1": build_info.commit_sha1,
"message": commit_message,
"commit_date": build_info.commit_timestamp,
"branch": build_info.branch,
"tag": build_info.describe.and_then(|describe| describe.as_tag()),
},
"machine_hostname": env.hostname,
"max_workloads": max_workloads,
"reason": reason
}))
.send()
.await
.context("sending invocation")?;
if !response.status().is_success() {
bail!(
"could not send new invocation: {}",
response.text().await.unwrap_or_else(|_| "unknown".into())
);
}
let invocation_uuid: Uuid =
response.json().await.context("could not deserialize invocation response as JSON")?;
Ok(invocation_uuid)
}
pub async fn create_workload(
dashboard_client: &Client,
invocation_uuid: Uuid,
workload: &Workload,
) -> anyhow::Result<Uuid> {
let response = dashboard_client
.put("workload")
.json(&json!({
"invocation_uuid": invocation_uuid,
"name": &workload.name,
"max_runs": workload.run_count,
}))
.send()
.await
.context("could not create new workload")?;
if !response.status().is_success() {
bail!("creating new workload failed: {}", response.text().await.unwrap())
let invocation_uuid: Uuid =
response.json().await.context("could not deserialize invocation response as JSON")?;
Ok(invocation_uuid)
}
let workload_uuid: Uuid =
response.json().await.context("could not deserialize JSON as UUID")?;
Ok(workload_uuid)
}
pub async fn create_workload(
&self,
invocation_uuid: Uuid,
workload: &Workload,
) -> anyhow::Result<Uuid> {
let Self::Client(dashboard_client) = self else { return Ok(Uuid::now_v7()) };
pub async fn create_run(
dashboard_client: Client,
workload_uuid: Uuid,
report: &BTreeMap<String, CallStats>,
) -> anyhow::Result<()> {
let response = dashboard_client
.put("run")
.json(&json!({
"workload_uuid": workload_uuid,
"data": report
}))
.send()
.await
.context("sending new run")?;
if !response.status().is_success() {
bail!(
"sending new run failed: {}",
response.text().await.unwrap_or_else(|_| "unknown".into())
)
let response = dashboard_client
.put("workload")
.json(&json!({
"invocation_uuid": invocation_uuid,
"name": &workload.name,
"max_runs": workload.run_count,
}))
.send()
.await
.context("could not create new workload")?;
if !response.status().is_success() {
bail!("creating new workload failed: {}", response.text().await.unwrap())
}
let workload_uuid: Uuid =
response.json().await.context("could not deserialize JSON as UUID")?;
Ok(workload_uuid)
}
pub async fn create_run(
&self,
workload_uuid: Uuid,
report: &BTreeMap<String, CallStats>,
) -> anyhow::Result<()> {
let Self::Client(dashboard_client) = self else { return Ok(()) };
let response = dashboard_client
.put("run")
.json(&json!({
"workload_uuid": workload_uuid,
"data": report
}))
.send()
.await
.context("sending new run")?;
if !response.status().is_success() {
bail!(
"sending new run failed: {}",
response.text().await.unwrap_or_else(|_| "unknown".into())
)
}
Ok(())
}
pub async fn cancel_on_ctrl_c(self, invocation_uuid: Uuid, abort_handle: AbortHandle) {
tracing::info!("press Ctrl-C to cancel the invocation");
match ctrl_c().await {
Ok(()) => {
tracing::info!(%invocation_uuid, "received Ctrl-C, cancelling invocation");
self.mark_as_failed(invocation_uuid, None).await;
abort_handle.abort();
}
Err(error) => tracing::warn!(
error = &error as &dyn std::error::Error,
"failed to listen to Ctrl-C signal, invocation won't be canceled on Ctrl-C"
),
}
}
pub async fn mark_as_failed(&self, invocation_uuid: Uuid, failure_reason: Option<String>) {
if let DashboardClient::Client(client) = self {
let response = client
.post("cancel-invocation")
.json(&json!({
"invocation_uuid": invocation_uuid,
"failure_reason": failure_reason,
}))
.send()
.await;
let response = match response {
Ok(response) => response,
Err(response_error) => {
tracing::error!(error = &response_error as &dyn std::error::Error, %invocation_uuid, "could not mark invocation as failed");
return;
}
};
if !response.status().is_success() {
tracing::error!(
%invocation_uuid,
"could not mark invocation as failed: {}",
response.text().await.unwrap()
);
return;
}
}
tracing::warn!(%invocation_uuid, "marked invocation as failed or canceled");
}
Ok(())
}

24
xtask/src/bench/mod.rs
View File

@ -50,6 +50,10 @@ pub struct BenchDeriveArgs {
#[arg(long, default_value_t = default_dashboard_url())]
dashboard_url: String,
/// Don't actually send results to the dashboard
#[arg(long)]
no_dashboard: bool,
/// Directory to output reports.
#[arg(long, default_value_t = default_report_folder())]
report_folder: String,
@ -103,11 +107,11 @@ pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
let assets_client =
Client::new(None, args.assets_key.as_deref(), Some(std::time::Duration::from_secs(3600)))?; // 1h
let dashboard_client = Client::new(
Some(format!("{}/api/v1", args.dashboard_url)),
args.api_key.as_deref(),
Some(std::time::Duration::from_secs(60)),
)?;
let dashboard_client = if args.no_dashboard {
dashboard::DashboardClient::new_dry()
} else {
dashboard::DashboardClient::new(&args.dashboard_url, args.api_key.as_deref())?
};
// reporting uses its own client because keeping the stream open to wait for entries
// blocks any other requests
@ -127,12 +131,12 @@ pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
// enter runtime
rt.block_on(async {
dashboard::send_machine_info(&dashboard_client, &env).await?;
dashboard_client.send_machine_info(&env).await?;
let commit_message = build_info.commit_msg.context("missing commit message")?.split('\n').next().unwrap();
let max_workloads = args.workload_file.len();
let reason: Option<&str> = args.reason.as_deref();
let invocation_uuid = dashboard::create_invocation(&dashboard_client, build_info, commit_message, env, max_workloads, reason).await?;
let invocation_uuid = dashboard_client.create_invocation( build_info, commit_message, env, max_workloads, reason).await?;
tracing::info!(workload_count = args.workload_file.len(), "handling workload files");
@ -167,7 +171,7 @@ pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
let abort_handle = workload_runs.abort_handle();
tokio::spawn({
let dashboard_client = dashboard_client.clone();
dashboard::cancel_on_ctrl_c(invocation_uuid, dashboard_client, abort_handle)
dashboard_client.cancel_on_ctrl_c(invocation_uuid, abort_handle)
});
// wait for the end of the main task, handle result
@ -178,7 +182,7 @@ pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
}
Ok(Err(error)) => {
tracing::error!(%invocation_uuid, error = %error, "invocation failed, attempting to report the failure to dashboard");
dashboard::mark_as_failed(dashboard_client, invocation_uuid, Some(error.to_string())).await;
dashboard_client.mark_as_failed(invocation_uuid, Some(error.to_string())).await;
tracing::warn!(%invocation_uuid, "invocation marked as failed following error");
Err(error)
},
@ -186,7 +190,7 @@ pub fn run(args: BenchDeriveArgs) -> anyhow::Result<()> {
match join_error.try_into_panic() {
Ok(panic) => {
tracing::error!("invocation panicked, attempting to report the failure to dashboard");
dashboard::mark_as_failed(dashboard_client, invocation_uuid, Some("Panicked".into())).await;
dashboard_client.mark_as_failed( invocation_uuid, Some("Panicked".into())).await;
std::panic::resume_unwind(panic)
}
Err(_) => {

16
xtask/src/bench/workload.rs
View File

@ -12,8 +12,9 @@ use uuid::Uuid;
use super::assets::Asset;
use super::client::Client;
use super::command::SyncMode;
use super::dashboard::DashboardClient;
use super::BenchDeriveArgs;
use crate::bench::{assets, dashboard, meili_process};
use crate::bench::{assets, meili_process};
#[derive(Deserialize)]
pub struct Workload {
@ -25,7 +26,7 @@ pub struct Workload {
}
async fn run_commands(
dashboard_client: &Client,
dashboard_client: &DashboardClient,
logs_client: &Client,
meili_client: &Client,
workload_uuid: Uuid,
@ -64,7 +65,7 @@ async fn run_commands(
#[tracing::instrument(skip(assets_client, dashboard_client, logs_client, meili_client, workload, master_key, args), fields(workload = workload.name))]
pub async fn execute(
assets_client: &Client,
dashboard_client: &Client,
dashboard_client: &DashboardClient,
logs_client: &Client,
meili_client: &Client,
invocation_uuid: Uuid,
@ -74,8 +75,7 @@ pub async fn execute(
) -> anyhow::Result<()> {
assets::fetch_assets(assets_client, &workload.assets, &args.asset_folder).await?;
let workload_uuid =
dashboard::create_workload(dashboard_client, invocation_uuid, &workload).await?;
let workload_uuid = dashboard_client.create_workload(invocation_uuid, &workload).await?;
let mut tasks = Vec::new();
@ -113,7 +113,7 @@ pub async fn execute(
#[allow(clippy::too_many_arguments)] // not best code quality, but this is a benchmark runner
#[tracing::instrument(skip(dashboard_client, logs_client, meili_client, workload, master_key, args), fields(workload = %workload.name))]
async fn execute_run(
dashboard_client: &Client,
dashboard_client: &DashboardClient,
logs_client: &Client,
meili_client: &Client,
workload_uuid: Uuid,
@ -202,7 +202,7 @@ async fn start_report(
}
async fn stop_report(
dashboard_client: &Client,
dashboard_client: &DashboardClient,
logs_client: &Client,
workload_uuid: Uuid,
filename: String,
@ -232,7 +232,7 @@ async fn stop_report(
.context("could not convert trace to report")?;
let context = || format!("writing report to {filename}");
dashboard::create_run(dashboard_client, workload_uuid, &report).await?;
dashboard_client.create_run(workload_uuid, &report).await?;
let mut output_file = std::io::BufWriter::new(
std::fs::File::options()