3821: Add normalized and detailed scores to documents returned by a query r=dureuill a=dureuill
# Pull Request
## Related issue
Fixes#3771
## What does this PR do?
### User standpoint
<details>
<summary>Request ranking score</summary>
```
echo '{
"q": "Badman dark knight returns",
"showRankingScore": true,
"limit": 10,
"attributesToRetrieve": ["title"]
}' | mieli search -i index-word-count-10-count
```
</details>
<details>
<summary>Response</summary>
```json
{
"hits": [
{
"title": "Batman: The Dark Knight Returns, Part 1",
"_rankingScore": 0.947520325203252
},
{
"title": "Batman: The Dark Knight Returns, Part 2",
"_rankingScore": 0.947520325203252
},
{
"title": "Batman Unmasked: The Psychology of the Dark Knight",
"_rankingScore": 0.6657594086021505
},
{
"title": "Legends of the Dark Knight: The History of Batman",
"_rankingScore": 0.6654905913978495
},
{
"title": "Angel and the Badman",
"_rankingScore": 0.2196969696969697
},
{
"title": "Angel and the Badman",
"_rankingScore": 0.2196969696969697
},
{
"title": "Batman",
"_rankingScore": 0.11553030303030302
},
{
"title": "Batman Begins",
"_rankingScore": 0.11553030303030302
},
{
"title": "Batman Returns",
"_rankingScore": 0.11553030303030302
},
{
"title": "Batman Forever",
"_rankingScore": 0.11553030303030302
}
],
"query": "Badman dark knight returns",
"processingTimeMs": 12,
"limit": 10,
"offset": 0,
"estimatedTotalHits": 46
}
```
</details>
- If adding a `showRankingScore` parameter to the search query, then documents returned by a search now contain an additional field `_rankingScore` that is a float bigger than 0 and lower or equal to 1.0. This field represents the relevancy of the document, relatively to the search query and the settings of the index, with 1.0 meaning "perfect match" and 0 meaning "not matching the query" (Meilisearch should never return documents not matching the query at all).
- The `sort` and `geosort` ranking rules do not influence the `_rankingScore`.
<details>
<summary>Request detailed ranking scores</summary>
```
echo '{
"q": "Badman dark knight returns",
"showRankingScoreDetails": true,
"limit": 5,
"attributesToRetrieve": ["title"]
}' | mieli search -i index-word-count-10-count
```
</details>
<details>
<summary>Response</summary>
```json
{
"hits": [
{
"title": "Batman: The Dark Knight Returns, Part 1",
"_rankingScoreDetails": {
"words": {
"order": 0,
"matchingWords": 4,
"maxMatchingWords": 4,
"score": 1.0
},
"typo": {
"order": 1,
"typoCount": 1,
"maxTypoCount": 4,
"score": 0.8
},
"proximity": {
"order": 2,
"score": 0.9545454545454546
},
"attribute": {
"order": 3,
"attributes_ranking_order": 1.0,
"attributes_query_word_order": 0.926829268292683,
"score": 0.926829268292683
},
"exactness": {
"order": 4,
"matchType": "noExactMatch",
"score": 0.26666666666666666
}
}
},
{
"title": "Batman: The Dark Knight Returns, Part 2",
"_rankingScoreDetails": {
"words": {
"order": 0,
"matchingWords": 4,
"maxMatchingWords": 4,
"score": 1.0
},
"typo": {
"order": 1,
"typoCount": 1,
"maxTypoCount": 4,
"score": 0.8
},
"proximity": {
"order": 2,
"score": 0.9545454545454546
},
"attribute": {
"order": 3,
"attributes_ranking_order": 1.0,
"attributes_query_word_order": 0.926829268292683,
"score": 0.926829268292683
},
"exactness": {
"order": 4,
"matchType": "noExactMatch",
"score": 0.26666666666666666
}
}
},
{
"title": "Batman Unmasked: The Psychology of the Dark Knight",
"_rankingScoreDetails": {
"words": {
"order": 0,
"matchingWords": 3,
"maxMatchingWords": 4,
"score": 0.75
},
"typo": {
"order": 1,
"typoCount": 1,
"maxTypoCount": 3,
"score": 0.75
},
"proximity": {
"order": 2,
"score": 0.6666666666666666
},
"attribute": {
"order": 3,
"attributes_ranking_order": 1.0,
"attributes_query_word_order": 0.8064516129032258,
"score": 0.8064516129032258
},
"exactness": {
"order": 4,
"matchType": "noExactMatch",
"score": 0.25
}
}
},
{
"title": "Legends of the Dark Knight: The History of Batman",
"_rankingScoreDetails": {
"words": {
"order": 0,
"matchingWords": 3,
"maxMatchingWords": 4,
"score": 0.75
},
"typo": {
"order": 1,
"typoCount": 1,
"maxTypoCount": 3,
"score": 0.75
},
"proximity": {
"order": 2,
"score": 0.6666666666666666
},
"attribute": {
"order": 3,
"attributes_ranking_order": 1.0,
"attributes_query_word_order": 0.7419354838709677,
"score": 0.7419354838709677
},
"exactness": {
"order": 4,
"matchType": "noExactMatch",
"score": 0.25
}
}
},
{
"title": "Angel and the Badman",
"_rankingScoreDetails": {
"words": {
"order": 0,
"matchingWords": 1,
"maxMatchingWords": 4,
"score": 0.25
},
"typo": {
"order": 1,
"typoCount": 0,
"maxTypoCount": 1,
"score": 1.0
},
"proximity": {
"order": 2,
"score": 1.0
},
"attribute": {
"order": 3,
"attributes_ranking_order": 1.0,
"attributes_query_word_order": 0.8181818181818182,
"score": 0.8181818181818182
},
"exactness": {
"order": 4,
"matchType": "noExactMatch",
"score": 0.3333333333333333
}
}
}
],
"query": "Badman dark knight returns",
"processingTimeMs": 9,
"limit": 5,
"offset": 0,
"estimatedTotalHits": 46
}
```
</details>
- If adding a `showRankingScoreDetails` parameter to the search query, then the returned documents will now contain an additional `_rankingScoreDetails` field that is a JSON object containing one field per ranking rule that was applied, whose value is a JSON object with the following fields:
- `order`: a number indicating the order this rule was applied (0 is the first applied ranking rule)
- `score` (except for `sort` and `geosort`): a float indicating how the document matched this particular rule.
- other fields that are specific to the rule, indicating for example how many words matched for a document and how many typos were counted in a matching document.
- If the `displayableAttributes` list is defined in the settings of the index, any ranking rule using an attribute **not** part of that list will be marked as `<hidden-rule>` in the `_rankingScoreDetails`.
- Search queries that are part of a `multi-search` requests are modified in the same way and each of the queries can take the `showRankingScore` and `showRankingScoreDetails` parameters independently. The results are still returned in separate lists and providing a unified list of results between multiple queries is not in the scope of this PR (but is unblocked by this PR and can be done manually by using the scores of the various documents).
### Implementation standpoint
- Fix difference in how the position of terms were computed at indexing time and query time: this difference meant that a query containing a hard separator would fail the exactness check.
- Fix the id reported by the sort ranking rule (very minor)
- Change how the cost of removing words is computed. After this change the cost no longer works for any other ranking rule than `words`. Also made `words` have a cost of 0 such that the entire cost of `words` is given by the termRemovalStrategy. The new cost computation makes it so the score is computed in a way consistent with the number of words in the query. Additionally, the words that appear in phrases in the query are also counted as matching words.
- When any score computation is requested through `showRankingScore` or `showRankingScoreDetails`, remove optimization where ranking rules are not executed on buckets of a single document: this is important to allow the computation of an accurate score.
- add virtual conditions to fid and position to always have the max cost: this ensures that the score is independent from the dataset
- the Position ranking rule now takes into account the distance to the position of the word in the query instead of the distance to the position 0.
- modified proximity ranking rule cost calculation so that the cost is 0 for documents that are perfectly matching the query
- Add a new `milli::score_details` module containing all the types that are involved in score computation.
- Make it so a bucket of result now contains a `ScoreDetails` and changed the ranking rules to produce their `ScoreDetails`.
- Expose the scores in the REST API.
- Add very light analytics for scoring.
- Update the search tests to add the expected scores.
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
3842: fix some typos r=dureuill a=cuishuang
# Pull Request
## Related issue
Fixes #<issue_number>
## What does this PR do?
- fix some typos
## 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: cui fliter <imcusg@gmail.com>
3670: Fix addition deletion bug r=irevoire a=irevoire
The first commit of this PR is a revert of https://github.com/meilisearch/meilisearch/pull/3667. It re-enable the auto-batching of addition and deletion of tasks. No new changes have been introduced outside of `milli`. So all the changes you see on the autobatcher have actually already been reviewed.
It fixes https://github.com/meilisearch/meilisearch/issues/3440.
### What was happening?
The issue was that the `external_documents_ids` generated in the `transform` were used in a very strange way that wasn’t compatible with the deletion of documents.
Instead of doing a clear merge between the external document IDs of the DB and the one returned by the transform + writing it on disk, we were doing some weird tricks with the soft-deleted to avoid writing the fst on disk as much as possible.
The new algorithm may be a bit slower but is way more straightforward and doesn’t change depending on if the soft deletion was used or not. Here is a list of the changes introduced:
1. We now do a clear distinction between the `new_external_documents_ids` coming from the transform and only held on RAM and the `external_documents_ids` coming from the DB.
2. The `new_external_documents_ids` (coming out of the transform) are now represented as an `fst`. We don't need to struggle with the hard, soft distinction + the soft_deleted => That's easier to understand
3. When indexing documents, we merge the `external_documents_ids` coming from the DB and the `new_external_documents_ids` coming from the transform.
### Other things introduced in this PR
Since we constantly have to write small, very specialized fuzzers for this kind of bug, we decided to push the one used to reproduce this bug.
It's not perfect, but it's easy to improve in the future.
It'll also run for as long as possible on every merge on the main branch.
Co-authored-by: Tamo <tamo@meilisearch.com>
Co-authored-by: Loïc Lecrenier <loic.lecrenier@icloud.com>
3835: Add more documentation to graph-based ranking rule algorithms + comment cleanup r=Kerollmops a=loiclec
In addition to documenting the `cheapest_path.rs` file, this PR cleans up a few outdated comments as well as some TODOs. These TODOs have been moved to https://github.com/meilisearch/meilisearch/issues/3776
Co-authored-by: Loïc Lecrenier <loic.lecrenier@icloud.com>
3768: Fix bugs in graph-based ranking rules + make `words` a graph-based ranking rule r=dureuill a=loiclec
This PR contains three changes:
## 1. Don't call the `words` ranking rule if the term matching strategy is `All`
This is because the purpose of `words` is only to remove nodes from the query graph. It would never do any useful work when the matching strategy was `All`. Remember that the universe was already computed before by computing all the docids corresponding to the "maximally reduced" query graph, which, in the case of `All`, is equal to the original graph.
## 2. The `words` ranking rule is replaced by a graph-based ranking rule.
This is for three reasons:
1. **performance**: graph-based ranking rules benefit from a lot of optimisations by default, which ensures that they are never too slow. The previous implementation of `words` could call `compute_query_graph_docids` many times if some words had to be removed from the query, which would be quite expensive. I was especially worried about its performance in cases where it is placed right after the `sort` ranking rule. Furthermore, `compute_query_graph_docids` would clone a lot of bitmaps many times unnecessarily.
2. **consistency**: every other ranking rule (except `sort`) is graph-based. It makes sense to implement `words` like that as well. It will automatically benefit from all the features, optimisations, and bug fixes that all the other ranking rules get.
3. **surfacing bugs**: as the first ranking rule to be called (most of the time), I'd like `words` to behave the same as the other ranking rules so that we can quickly detect bugs in our graph algorithms. This actually already happened, which is why this PR also contains a bug fix.
## 3. Fix the `update_all_costs_before_nodes` function
It is a bit difficult to explain what was wrong, but I'll try. The bug happened when we had graphs like:
<img width="730" alt="Screenshot 2023-05-16 at 10 58 57" src="https://github.com/meilisearch/meilisearch/assets/6040237/40db1a68-d852-4e89-99d5-0d65757242a7">
and we gave the node `is` as argument.
Then, we'd walk backwards from the node breadth-first. We'd update the costs of:
1. `sun`
2. `thesun`
3. `start`
4. `the`
which is an incorrect order. The correct order is:
1. `sun`
2. `thesun`
3. `the`
4. `start`
That is, we can only update the cost of a node when all of its successors have either already been visited or were not affected by the update to the node passed as argument. To solve this bug, I factored out the graph-traversal logic into a `traverse_breadth_first_backward` function.
Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
Co-authored-by: Louis Dureuil <louis@meilisearch.com>
3757: Adjust the cost of edges in the `position` ranking rule by bucketing positions more aggressively r=loiclec a=loiclec
This PR significantly improves the performance of the `position` ranking rule when:
1. a query contains many words
2. the `position` ranking rule needs to be called many times
3. the score of the documents according to `position` is high
These conditions greatly increase:
1. the number of edge traversals that are needed to find a valid path from the `start` node to the `end` node
2. the number of edges that need to be deleted from the graph, and therefore the number of times that we need to recompute all the possible costs from START to END
As a result, a majority of the search time is spent in `visit_condition`, `visit_node`, and `update_all_costs_before_node`. This is frustrating because it often happens when the "universe" given to the rule consists of only a handful of document ids.
By limiting the number of possible edges between two nodes from `20` to `10`, we:
1. reduce the number of possible costs from START to END
2. reduce the number of edges that will be deleted
3. make it faster to update the costs after deleting an edge
4. reduce the number of buckets that need to be computed
In terms of relevancy, I don't think we lose or gain much. We still prefer terms that are in a lower positions, with decreasing precision as we go further. The previous choice of bucketing wasn't chosen in a principled way, and neither is this one. They both "feel" right to me.
Co-authored-by: Loïc Lecrenier <loic.lecrenier@me.com>
Co-authored-by: meili-bors[bot] <89034592+meili-bors[bot]@users.noreply.github.com>
