514: Stop flattening every field r=Kerollmops a=irevoire
When we need to flatten a document:
* The primary key contains a `.`.
* Some fields need to be flattened
Instead of flattening the whole object and thus creating a lot of allocations with the `serde_json_flatten_crate`, we instead generate a minimal sub-object containing only the fields that need to be flattened.
That should create fewer allocations and thus index faster.
---------
```
group indexing_main_e1e362fa indexing_stop-flattening-every-field_40d1bd6b
----- ---------------------- ---------------------------------------------
indexing/Indexing geo_point 1.99 23.7±0.23s ? ?/sec 1.00 11.9±0.21s ? ?/sec
indexing/Indexing movies in three batches 1.00 18.2±0.24s ? ?/sec 1.01 18.3±0.29s ? ?/sec
indexing/Indexing movies with default settings 1.00 17.5±0.09s ? ?/sec 1.01 17.7±0.26s ? ?/sec
indexing/Indexing songs in three batches with default settings 1.00 64.8±0.47s ? ?/sec 1.00 65.1±0.49s ? ?/sec
indexing/Indexing songs with default settings 1.00 54.9±0.99s ? ?/sec 1.01 55.7±1.34s ? ?/sec
indexing/Indexing songs without any facets 1.00 50.6±0.62s ? ?/sec 1.01 50.9±1.05s ? ?/sec
indexing/Indexing songs without faceted numbers 1.00 54.0±1.14s ? ?/sec 1.01 54.7±1.13s ? ?/sec
indexing/Indexing wiki 1.00 996.2±8.54s ? ?/sec 1.02 1021.1±30.63s ? ?/sec
indexing/Indexing wiki in three batches 1.00 1136.8±9.72s ? ?/sec 1.00 1138.6±6.59s ? ?/sec
```
So basically everything slowed down a liiiiiittle bit except the dataset with a nested field which got twice faster
Co-authored-by: Tamo <tamo@meilisearch.com>
505: normalize exact words r=curquiza a=MarinPostma
Normalize the exact words, as specified in the specification.
Co-authored-by: ad hoc <postma.marin@protonmail.com>
483: Enhance matching words r=Kerollmops a=ManyTheFish
# Summary
Enhance milli word-matcher making it handle match computing and cropping.
# Implementation
## Computing best matches for cropping
Before we were considering that the first match of the attribute was the best one, this was accurate when only one word was searched but was missing the target when more than one word was searched.
Now we are searching for the best matches interval to crop around, the chosen interval is the one:
1) that have the highest count of unique matches
> for example, if we have a query `split the world`, then the interval `the split the split the` has 5 matches but only 2 unique matches (1 for `split` and 1 for `the`) where the interval `split of the world` has 3 matches and 3 unique matches. So the interval `split of the world` is considered better.
2) that have the minimum distance between matches
> for example, if we have a query `split the world`, then the interval `split of the world` has a distance of 3 (2 between `split` and `the`, and 1 between `the` and `world`) where the interval `split the world` has a distance of 2. So the interval `split the world` is considered better.
3) that have the highest count of ordered matches
> for example, if we have a query `split the world`, then the interval `the world split` has 2 ordered words where the interval `split the world` has 3. So the interval `split the world` is considered better.
## Cropping around the best matches interval
Before we were cropping around the interval without checking the context.
Now we are cropping around words in the same context as matching words.
This means that we will keep words that are farther from the matching words but are in the same phrase, than words that are nearer but separated by a dot.
> For instance, for the matching word `Split` the text:
`Natalie risk her future. Split The World is a book written by Emily Henry. I never read it.`
will be cropped like:
`…. Split The World is a book written by Emily Henry. …`
and not like:
`Natalie risk her future. Split The World is a book …`
Co-authored-by: ManyTheFish <many@meilisearch.com>
We need to store all the external id (primary key) in a hashmap
associated to their internal id during.
The smartstring remove heap allocation / memory usage and should
improve the cache locality.
486: Update version (v0.25.0) r=curquiza a=curquiza
v0.25.0 will be released once #478 is merged
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
472: Remove useless variables in proximity r=Kerollmops a=ManyTheFish
Was passing by plane sweep algorithm to find some inspiration, and I discover that we have useless variables that were not detected because of the recursive function.
Co-authored-by: ManyTheFish <many@meilisearch.com>
466: Bump version to 0.23.1 r=curquiza a=Kerollmops
This PR bumps the crate versions to 0.23.1. Nothing seems to be breaking in the next release.
Co-authored-by: Kerollmops <clement@meilisearch.com>
467: optimize prefix database r=Kerollmops a=MarinPostma
This pr introduces two optimizations that greatly improve the speed of computing prefix databases.
- The time that it takes to create the prefix FST has been divided by 5 by inverting the way we iterated over the words FST.
- We unconditionally and needlessly checked for documents to remove in `word_prefix_pair`, which caused an iteration over the whole database.
Co-authored-by: ad hoc <postma.marin@protonmail.com>
> "Attribute `{}` is not sortable. This index doesn't have configured sortable attributes."
> "Attribute `{}` is not sortable. Available sortable attributes are: `{}`."
coexist in the error handling
436: Speed up the word prefix databases computation time r=Kerollmops a=Kerollmops
This PR depends on the fixes done in #431 and must be merged after it.
In this PR we will bring the `WordPrefixPairProximityDocids`, `WordPrefixDocids` and, `WordPrefixPositionDocids` update structures to a new era, a better era, where computing the word prefix pair proximities costs much fewer CPU cycles, an era where this update structure can use the, previously computed, set of new word docids from the newly indexed batch of documents.
---
The `WordPrefixPairProximityDocids` is an update structure, which means that it is an object that we feed with some parameters and which modifies the LMDB database of an index when asked for. This structure specifically computes the list of word prefix pair proximities, which correspond to a list of pairs of words associated with a proximity (the distance between both words) where the second word is not a word but a prefix e.g. `s`, `se`, `a`. This word prefix pair proximity is associated with the list of documents ids which contains the pair of words and prefix at the given proximity.
The origin of the performances issue that this struct brings is related to the fact that it starts its job from the beginning, it clears the LMDB database before rewriting everything from scratch, using the other LMDB databases to achieve that. I hope you understand that this is absolutely not an optimized way of doing things.
Co-authored-by: Clément Renault <clement@meilisearch.com>
Co-authored-by: Kerollmops <clement@meilisearch.com>
442: fix phrase search r=curquiza a=MarinPostma
Run the exact match search on 7 words windows instead of only two. This makes false positive very very unlikely, and impossible on phrase query that are less than seven words.
Co-authored-by: ad hoc <postma.marin@protonmail.com>
431: Fix and improve word prefix pair proximity r=ManyTheFish a=Kerollmops
This PR first fixes the algorithm we used to select and compute the word prefix pair proximity database. The previous version was skipping nearly all of the prefixes. The issue is that this fix made this method to take more time and we were trying to reduce the time spent in it.
With `@ManyTheFish` we found out that we could skip some of the work we were doing by:
- discarding the prefixes that were shorter than a specific threshold (default: 2).
- discarding the word prefix pairs with proximity bigger than a specific threshold (default: 4).
- remove the unused threshold that was specifying a minimum amount of word docids to merge.
We will take more time to do some more optimization, like stop clearing and recomputing from scratch the database, we will compute the subsets of keys to create, keep and merge. This change is a little bit more complex than what this PR does.
I keep this PR as a draft as I want to further test the real gain if it is enough or not if it is valid or not. I advise reviewers to review commit by commit to see the changes bit by bit, reviewing the whole PR can be hard.
Co-authored-by: Clément Renault <clement@meilisearch.com>
433: fix(filter): Fix two bugs. r=Kerollmops a=irevoire
- Stop lowercasing the field when looking in the field id map
- When a field id does not exist it means there is currently zero
documents containing this field thus we return an empty RoaringBitmap
instead of throwing an internal error
Will fix https://github.com/meilisearch/MeiliSearch/issues/2082 once meilisearch is released
Co-authored-by: Tamo <tamo@meilisearch.com>
426: Fix search highlight for non-unicode chars r=ManyTheFish a=Samyak2
# Pull Request
## What does this PR do?
Fixes https://github.com/meilisearch/MeiliSearch/issues/1480
<!-- Please link the issue you're trying to fix with this PR, if none then please create an issue first. -->
## PR checklist
Please check if your PR fulfills the following requirements:
- [x] Does this PR fix an existing issue?
- [x] Have you read the contributing guidelines?
- [x] Have you made sure that the title is accurate and descriptive of the changes?
## Changes
The `matching_bytes` function takes a `&Token` now and:
- gets the number of bytes to highlight (unchanged).
- uses `Token.num_graphemes_from_bytes` to get the number of grapheme clusters to highlight.
In essence, the `matching_bytes` function now returns the number of matching grapheme clusters instead of bytes.
Added proper highlighting in the HTTP UI:
- requires dependency on `unicode-segmentation` to extract grapheme clusters from tokens
- `<mark>` tag is put around only the matched part
- before this change, the entire word was highlighted even if only a part of it matched
## Questions
Since `matching_bytes` does not return number of bytes but grapheme clusters, should it be renamed to something like `matching_chars` or `matching_graphemes`? Will this break the API?
Thank you very much `@ManyTheFish` for helping 😄
Co-authored-by: Samyak S Sarnayak <samyak201@gmail.com>
- Stop lowercasing the field when looking in the field id map
- When a field id does not exist it means there is currently zero
documents containing this field thus we returns an empty RoaringBitmap
instead of throwing an internal error
The `matching_bytes` function takes a `&Token` now and:
- gets the number of bytes to highlight (unchanged).
- uses `Token.num_graphemes_from_bytes` to get the number of grapheme
clusters to highlight.
In essence, the `matching_bytes` function returns the number of matching
grapheme clusters instead of bytes. Should this function be renamed
then?
Added proper highlighting in the HTTP UI:
- requires dependency on `unicode-segmentation` to extract grapheme
clusters from tokens
- `<mark>` tag is put around only the matched part
- before this change, the entire word was highlighted even if only a
part of it matched
returned metaimprove document addition returned metaimprove document
addition returned metaimprove document addition returned metaimprove
document addition returned metaimprove document addition returned
metaimprove document addition returned meta
407: Update version for the next release (v0.20.0) r=curquiza a=curquiza
Breaking because of #405 and #406
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
402: Optimize document transform r=MarinPostma a=MarinPostma
This pr optimizes the transform of documents additions in the obkv format. Instead on accepting any serializable objects, we instead treat json and CSV specifically:
- For json, we build a serde `Visitor`, that transform the json straight into obkv without intermediate representation.
- For csv, we directly write the lines in the obkv, applying other optimization as well.
Co-authored-by: marin postma <postma.marin@protonmail.com>
390: Add helper methods on the settings r=Kerollmops a=irevoire
This would be a good addition to look at the content of a setting without consuming it.
It’s useful for analytics.
Co-authored-by: Irevoire <tamo@meilisearch.com>
384: Replace memmap with memmap2 r=Kerollmops a=palfrey
[memmap is unmaintained](https://rustsec.org/advisories/RUSTSEC-2020-0077.html) and needs replacing. memmap2 is a drop-in replacement fork that's well maintained. Note that the version numbers got reset on fork, hence the lower values.
Co-authored-by: Tom Parker-Shemilt <palfrey@tevp.net>
388: fix primary key inference r=MarinPostma a=MarinPostma
The primary key is was infered from a hashtable index of the field. For this reason the order in which the fields were interated upon was not deterministic, and the primary key was chosed ffrom the first field containing "id".
This fix sorts the the index by field_id when infering the primary key.
Co-authored-by: mpostma <postma.marin@protonmail.com>
Instead of using an arbitrary limit we encode the absolute position in a u32
using one strong u16 for the field id and a weak u16 for the relative position in the attribute.
386: fix obkv document r=curquiza a=MarinPostma
When serializing a document, the serializer resolved the field_id of the current field and immediately added it to the obkv document under construction. The issue with that is that obkv expects the fields to be inserted in order, and when a document with out of order fields was added, obkv failed to insert the field.
The current fix first resolves each field_id, and adds all the fields to a temporary `BTreeMap`, until `end` is called on the map serializer, where all the fields are added to the obkv at once, and in order.
Co-authored-by: mpostma <postma.marin@protonmail.com>
Latitude are not supposed to go beyound 90 degrees or below -90.
The same goes for longitude with 180 or -180.
This was badly implemented in the filters, and was not implemented for the AscDesc rules.
379: Revert "Change chunk size to 4MiB to fit more the end user usage" r=curquiza a=ManyTheFish
Reverts meilisearch/milli#370
Co-authored-by: Many <legendre.maxime.isn@gmail.com>
376: Stop casting integer docids to string r=Kerollmops a=irevoire
When a docid is an integer, we stop casting it to a string, and thus we don't add `"` around it.
Co-authored-by: Tamo <tamo@meilisearch.com>
373: Improve error message for bad sort syntax with geosearch r=Kerollmops a=irevoire
`@Kerollmops` This should be the last PR for the geosearch and error handling, sorry for doing it in so many steps 😬
Co-authored-by: Tamo <tamo@meilisearch.com>
372: Fix Meilisearch 1714 r=Kerollmops a=ManyTheFish
The bug comes from the typo tolerance, to know how many typos are accepted we were counting bytes instead of characters in a word.
On Chinese Script characters, we were allowing 2 typos on 3 characters words.
We are now counting the number of char instead of counting bytes to assign the typo tolerance.
Related to [Meilisearch#1714](https://github.com/meilisearch/MeiliSearch/issues/1714)
Co-authored-by: many <maxime@meilisearch.com>
360: Update version for the next release (v0.14.0) r=Kerollmops a=curquiza
Release containing the geosearch, cf #322
Co-authored-by: Clémentine Urquizar <clementine@meilisearch.com>
322: Geosearch r=ManyTheFish a=irevoire
This PR introduces [basic geo-search functionalities](https://github.com/meilisearch/specifications/pull/59), it makes the engine able to index, filter and, sort by geo-point. We decided to use [the rstar library](https://docs.rs/rstar) and to save the points in [an RTree](https://docs.rs/rstar/0.9.1/rstar/struct.RTree.html) that we de/serialize in the index database [by using serde](https://serde.rs/) with [bincode](https://docs.rs/bincode). This is not an efficient way to query this tree as it will consume a lot of CPU and memory when a search is made, but at least it is an easy first way to do so.
### What we will have to do on the indexing part:
- [x] Index the `_geo` fields from the documents.
- [x] Create a new module with an extractor in the `extract` module that takes the `obkv_documents` and retrieves the latitude and longitude coordinates, outputting them in a `grenad::Reader` for further process.
- [x] Call the extractor in the `extract::extract_documents_data` function and send the result to the `TypedChunk` module.
- [x] Get the `grenad::Reader` in the `typed_chunk::write_typed_chunk_into_index` function and store all the points in the `rtree`
- [x] Delete the documents from the `RTree` when deleting documents from the database. All this can be done in the `delete_documents.rs` file by getting the data structure and removing the points from it, inserting it back after the modification.
- [x] Clearing the `RTree` entirely when we clear the documents from the database, everything happens in the `clear_documents.rs` file.
- [x] save a Roaring bitmap of all documents containing the `_geo` field
### What we will have to do on the query part:
- [x] Filter the documents at a certain distance around a point, this is done by [collecting the documents from the searched point](https://docs.rs/rstar/0.9.1/rstar/struct.RTree.html#method.nearest_neighbor_iter) while they are in range.
- [x] We must introduce new `geoLowerThan` and `geoGreaterThan` variants to the `Operator` filter enum.
- [x] Implement the `negative` method on both variants where the `geoGreaterThan` variant is implemented by executing the `geoLowerThan` and removing the results found from the whole list of geo faceted documents.
- [x] Add the `_geoRadius` function in the pest parser.
- [x] Introduce a `_geo` ascending ranking function that takes a point in parameter, ~~this function must keep the iterator on the `RTree` and make it peekable~~ This was not possible for now, we had to collect the whole iterator. Only the documents that are part of the candidates must be sent too!
- [x] This ascending ranking rule will only be active if the search is set up with the `_geoPoint` parameter that indicates the center point of the ascending ranking rule.
-----------
- On Meilisearch part: We must introduce a new concept, returning the documents with a new `_geoDistance` field when it passed by the `_geo` ranking rule, this has never been done before. We could maybe just do it afterward when the documents have been retrieved from the database, computing the distance from the `_geoPoint` and all of the documents to be returned.
Co-authored-by: Irevoire <tamo@meilisearch.com>
Co-authored-by: cvermand <33010418+bidoubiwa@users.noreply.github.com>
Co-authored-by: Tamo <tamo@meilisearch.com>
342: Let the caller decide what kind of error they want to returns when parsing `AscDesc` r=Kerollmops a=irevoire
This is one possible fix for #339
We would then need to patch these lines https://github.com/meilisearch/MeiliSearch/blob/main/meilisearch-http/src/index/search.rs#L110-L114 to return the error we want.
Another solution would be to add a parameter to the `from_str` to specify which context we are in.
Co-authored-by: Tamo <tamo@meilisearch.com>
344: Move the sort ranking rule before the exactness ranking rule r=ManyTheFish a=Kerollmops
This PR moves the sort ranking rule at the 5th position by default, right before the exactness one.
Co-authored-by: Kerollmops <clement@meilisearch.com>
341: Throw a query time error when a sort parameter is used but the sort ranking rule is missing r=Kerollmops a=Kerollmops
This PR makes the engine throw an error for when the ranking rules don't contain the `sort` rule, the `sortable_fields` are correctly set but the user tries to use the `sort` query parameter. Doing so will have no effect on the returned documents so we preferred returning an error to help debug this.
That's breaking on the MeiliSearch side as we added a new variant to the `UserError` enum.
Co-authored-by: Kerollmops <clement@meilisearch.com>
308: Implement a better parallel indexer r=Kerollmops a=ManyTheFish
Rewrite the indexer:
- enhance memory consumption control
- optimize parallelism using rayon and crossbeam channel
- factorize the different parts and make new DB implementation easier
- optimize and fix prefix databases
Co-authored-by: many <maxime@meilisearch.com>
