4275: Flatten settings r=dureuill a=dureuill
# Pull Request
## Related issue
Initial internal feedback seems to indicate that the current shape of the `embedders` setting is undesirable: it has too much depth.
This PR changes this by flattening the structure of the embedders to the following:
```json5
// NEW structure
"embedders": {
// still starts with the embedder name
"default": {
"source": "huggingFace", // now a string
// properties of the source are all at the same level as the source
"model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
"revision": "a9c555277f9bcf24f28fa5e092e665fc6f7c49cd",
"documentTemplate": "A product titled '{{doc.title}}'" // now a string
}
}
```
By comparison, the old structure was:
```json5
// PREVIOUS version, no longer working with this PR
"embedders": {
// still starts with the embedder name
"default": {
"source": {
"huggingFace": {
"model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
"revision": "a9c555277f9bcf24f28fa5e092e665fc6f7c49cd"
},
"documentTemplate": {
"template": "A product titled '{{doc.title}}'" // now a string
}
}
}
```
The fields that are accepted in the new version of the `embedders` setting are depending on the value of the `source` field:
```json5
// huggingFace
"embedders": {
"default": {
"source": "huggingFace",
"model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
"revision": "a9c555277f9bcf24f28fa5e092e665fc6f7c49cd",
"documentTemplate": "A product titled '{{doc.title}}'"
}
}
// openAi
"embedders": {
"default": {
"source": "openAi",
"model": "text-embedding-ada-002",
"apiKey": "open_ai_api_key",
"documentTemplate": "A product titled '{{doc.title}}'"
}
}
// userProvided
"embedders": {
"default": {
"source": "userProvided",
"dimensions": 42, // mandatory
}
}
```
## What does this PR do?
- Flatten the settings structure
- Validate the prompt earlier to return a synchronous error on setting change rather than in the failing task
- Make it an error to pass a field for the wrong source (see above for allowed fields for each source)
- Not changed: It is still an error not to pass `dimensions` to the `userProvided` embedder
- If `source` was specified in the settings, validate the setting early to return a synchronous error in case of a missing mandatory field for the userProvided source (dimensions) or a forbidden field for the specified source.
- If `source` was not specified in the settings, still validate the setting, but only at indexing time, by using the source stored in the DB.
- Resets all values if the source changes, even if the user did not reset them explicitly.
## PR checklist
Please check if your PR fulfills the following requirements:
- [ ] Change the public facing guide for using the API
- [ ] Change examples of use in the changelog
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
4272: Don't pass default revision when the model is explicitly set in config r=Kerollmops a=dureuill
# Pull Request
## Related issue
Fixes#4271
## What does this PR do?
- When the `model` is explicitly set in the `embedders` setting, we reset the `revision` to `None`, such that if the user doesn't specify a revision, the head of the model repository is chosen.
- Not changed: If the user specifies a revision, it applies, like previously.
- Not changed: If the user doesn't specify a model, the default model with the default revision applies, like previously.
## Manual testing on a fresh DB
1. Enable experimental feature:
```sh
curl \
-X PATCH 'http://localhost:7700/experimental-features/' \
-H 'Content-Type: application/json' -H 'Authorization: Bearer foo' \
--data-binary '{ "vectorStore": true
}'
```
2. Send settings with a specified model but no specified revision:
```sh
curl \
-X PATCH 'http://localhost:7700/indexes/products/settings' \
-H 'Content-Type: application/json' --data-binary \
'{ "embedders": { "default": { "source": { "huggingFace": { "model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2" } }, "documentTemplate": { "template": "A product titled '{{doc.title}}'"} } } }'
```
3. Check that the task was successful:
```sh
curl 'http://localhost:7700/tasks/0'
{"uid":0,"indexUid":"products","status":"succeeded","type":"settingsUpdate","canceledBy":null,"details":{"embedders":{"default":{"source":{"huggingFace":{"model":"sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"}},"documentTemplate":{"template":"A product titled {{doc.title}}"}}}},"error":null,"duration":"PT0.001892S","enqueuedAt":"2023-12-20T09:17:01.73789Z","startedAt":"2023-12-20T09:17:01.73854Z","finishedAt":"2023-12-20T09:17:01.740432Z"}
```
4. Send documents to index:
```sh
curl 'https://localhost:7700/indexes/products/documents' -H 'Content-Type: application/json' --data-binary '{"id": 0, "title": "Best product"}'
```
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
4269: Remove dependency that requires libstdc++ r=dureuill a=dureuill
Removes the dependency that caused the additional runtime dependency on libstdc++ by disabling the default features of the hf tokenizer.
## Discussion
- This removes a feature that is using a C++ dependency and is supposed to accelerate the tokenizer. As the tokenizer is likely to be a significant bottleneck for embedding texts using a HF model, this is an issue.
- We should at least rerun the movies vector indexing and check that it still works correctly and that it has a runtime in the ballpark of what it used to be.
Co-authored-by: Louis Dureuil <louis.dureuil@xinra.net>
4268: Add libstdc++ in Dockerfile r=curquiza a=sanders41
# Pull Request
## Related issue
Fixes#4267
## What does this PR do?
- Add libstdc++ in the Dockerfile
## PR checklist
Please check if your PR fulfills the following requirements:
- [x] Does this PR fix an existing issue, or have you listed the changes applied in the PR description (and why they are needed)?
- [x] Have you read the contributing guidelines?
- [x] Have you made sure that the title is accurate and descriptive of the changes?
Thank you so much for contributing to Meilisearch!
Co-authored-by: Paul Sanders <psanders1@gmail.com>
4262: Update version for the next release (v1.6.0) in Cargo.toml r=curquiza a=meili-bot
⚠️ This PR is automatically generated. Check the new version is the expected one and Cargo.lock has been updated before merging.
Co-authored-by: curquiza <curquiza@users.noreply.github.com>
4257: Change proximity precision settings r=dureuill a=ManyTheFish
- [x] Add proximity_precision value into the analytics
- [x] Change the naming of `attributeScale` and `wordScale` into `byAttribute` and `byWord`
- [x] Remove proximityPrecision from the experimental feature
Co-authored-by: ManyTheFish <many@meilisearch.com>
Co-authored-by: Many the fish <many@meilisearch.com>
4226: Hybrid search r=dureuill a=dureuill
Allows to perform hybrid search requests that combine the results of semantic and keyword search and automatically generate embeddings.
## How to use
See [feature description](https://meilisearch.notion.site/v1-6-Hybrid-Search-Embedders-ea42c82f90cc4bc0be1eeb917c1118c8)
## Changes
- work is based on #4213
- milli::new search now takes an input universe directly, rather than computing it from a filter. This adds flexibility to require results on a subset of documents
- vector search is now a regular ranking rule (akin to sort and geosort) and reports its score as a ScoreDetail
- separate keyword search and vector search functions, vector search now respects (geo)sort ranking rules
- add automatic embedding
- add hybrid search
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
Co-authored-by: ManyTheFish <many@meilisearch.com>
