hi everyone! is it possible that, when using hybrid search, I've a facet that counts 80 results and when I add it as a filter I get 120 results? It's very strange!

With a filter added, the actual tokens that match the query for keyword search can change and this can fetch a different set of results. Take a look at

max_candidates

.

Thank you for your answer! It seems that the parameter doesn't affect the search and the facets results! I always have more results with active filter than on the facet without the active filter!

Does the same happen with only keyword search or only vector search?

only when using vector search both with hybrid search

query_by: 'embedding,title,content',
vector_query: 'embedding:([], alpha: 0.5, distance_threshold: 0.8, k: 200)',

and with vector only

query_by: 'embedding',
vector_query: 'embedding:([], alpha: 1, distance_threshold: 2, k: 200)',

I suspect that when filtering is added, the approximate knn search on the graph takes a different search path which unearths more candidates.

Oh yes I agree with you! I think that this behaviour may occur beacuse filtering is done before the knn search so that knn is done on a different dataset. I think that filtering has to be done after the knn search to remove unmatched documents. Is it possible?

Filtering is done along with the knn search. But when search takes a wider path, more candidates could be picked because a larger portion of the graph is traversed. You can try setting a larger M value for the hnsw index construction to mitigate this.