Is there a way to limit Hybrid results by score? (Similar to “near vector” with certainty)?
I have 2 words in the database (vectorized): 1). “Airplanes” 2). “Helicopters”
I want to do a search about “sandwiches” and return nothing using hybrid. Currently I am unable to do that and I just get the “closest” results even though they are very far. autolimit did not work for me.
Thanks
Hi @msj242,
Currently, it is not possible to only return results with some level of a match.
The main challenge is that the score coming back from keyword search is relative, and it is hard to assign it a value where score x is good or bad, as the calculated score is based on a number of keywords and number of objects in the database.
But I think there might be something we could do in Weaviate to eliminate really poor results, like in your case.
I’ve created a GitHub issue, which proposes introduction of acceptance threshold, which would allow us to filter out results that provide poor results on both fronts (keyword and vector search).
Hi @msj242 !
I am not sure you can limit those directly from the query, as it will combine the vector search and keyword.
Autolimit indeed will probably not help here, as it will still give you the vector search results that, on the given case, is closer but not related to you query.
Maybe an approach is to do keyword search + rerank?
