Hi, I find non-consistency in details to be surprising. I always use the same hybrid search with the same settings, but with some queries it returns usual ranking rules and in some of them the vectorSort only.

As I understood from your response, it is intended behavior and there is a certain mechanism under the hood of hybrid search, that uses the semanticRatio score from provided request and outputs the details accordingly if vectorSort was really high in it's ratio?

Under the hood, hybrid search performs two search requests: full-text search and semantic search.

For a hybrid search with limit = 20 and offset = 0, the algorithm is roughly the following:

1. perform full-text search, which returns up to 20 results whose ranking score details contain the full list of ranking rules.
2. perform semantic search, which returns 20 results whose ranking score details contain vectorSort instead of relevancy rules.
3. compare the ranking score details between the documents in the two lists of results. This is done by aggregating the scores given by contiguous relevancy rules into a single relevancy score, and comparing it with the similarity score of vectorSort. This is where the semanticRatio comes into play. When comparing the relevancy score to the similarity score, we multiply the relevancyScore by 1 - semanticRatio, and the semantic score by the semantic ratio. Any sort ranking rule is compared depending on its direction and value. The document with the highest score is added to the merged list of results
4. Return the merged list of 20 results.

To give an example, with semanticRatio = 0.6, limit = 4, offset = 0, imagine we have the full-text search return:

1. foo with relevancy score = 0.99 => weighted score = 0.99 * (1 - 0.6) = 0.99 * 0.4 = 0.396
2. bar with relevancy score = 0.85 => weighted score = 0.34
3. baz with relevancy score = 0.75 => weighted score = 0.30
4. foobar with relevancy score = 0.20 => weighted score = 0.08
   On the other hand, the semantic search returns:
5. fizz with similarity score = 0.74 => weighted score = 0.74 * 0.6 = 0.44
6. baz with similarity score = 0.62 => weighted score = 0.37
7. foobaz with similarity score = 0.59 => weighted score = 0.35
8. footer with similarity score = 0.55 => weighted score = 0.33

Then the merged list of results will be:

1. fizz with similarity score = 0.74 => weighted score = 0.74 * 0.6 = 0.44
2. foo with relevancy score = 0.99 => weighted score = 0.99 * (1 - 0.6) = 0.99 * 0.4 = 0.396
3. baz with similarity score = 0.62 => weighted score = 0.37
4. foobaz with similarity score = 0.59 => weighted score = 0.35

From this conclusion, I would imply a question: will it be possible to get around the meilisearch to get the keyword search ranking rules details even if semantic search was engaged? I use this analytics to improve my pages content for better result.
Anyway, if it's not feasible in near future, probably will make some debugging points in source code to send it to my logging system.

will it be possible to get around the meilisearch to get the keyword search ranking rules details even if semantic search was engaged?

No, it is not. The documents that are returned by the semantic search are not necessarily results from the keyword search.

Implementing this would entail instructing Meilisearch to do a keyword search restricted to the set of documents returned by the vector search, which is a large cost for the use case of analytics.

Your use case sounds very interesting though, I'm curious, how are you using the ranking score details to compute the analytics? Would it be possible to use the agregated ranking score (showRankingScore: true, then _rankingScore in the returned document) instead?

Sorry if bothered you and thank you for helping me.

You're welcome, it is my pleasure ☺️

Implementing this would entail instructing Meilisearch to do a keyword search restricted to the set of documents returned by the vector search, which is a large cost for the use case of analytics.

Maybe I didn't understand you, but what if, knowing that keyword search is performed earlier than vector search, why not keep the details for the documents that had the biggest vectorSort value? So not omitting the details with the new vector details, but having the full details together.

how are you using the ranking score details to compute the analytics?

I noticed that changing ranking rules is very crucial for accuracy in my dataset, so I decided to watch the behaviour of meilisearch and maybe even optimize ranking rules further. So I watch why the document was chosen and what can I do to make it lower or higher in the result list. And based on the details I optimize ranking rules themselves or optimize the content inside of documents so they match the queries they intended to match. Maybe I am a perfectionist 😄

Would it be possible to use the agregated ranking score (showRankingScore: true, then _rankingScore in the returned document) instead?

I also use that. But I have noticed that this value is not an absolute, it is relative to the whole dataset. By this I mean that when some document score evaluates to 0.66, it is 0.66 based on other documents in the database, and so this value may increase or decrease based on what other documents I have in the database and how they match the search query.
I see the _rankingScoreDetails to be more useful when it comes to optimizing the settings.
I am not talking about it in a bad sense, and maybe the current state of it is the only solution, I still have to learn a lot about search engines and ranking mechanism 😄

BTW here is my other discussion that slightly touches the topic of _rankingScore: https://github.com/orgs/meilisearch/discussions/788 - this is just my thoughts aloud