Harrison/faiss score #6341
Harrison/faiss score #6341
Conversation
Allow FAISS's similarity_search_with_score_by_vector to accept kwargs,
specifically: `score_threshold`.
Fixes error for document compression users in cases where there are no
relevant docs - for instance:
```
human_message: "hi" (<-- will return no relevant docs depending on score_threshold)
> IndexError: index 0 is out of bounds for axis 0 with size 0
```
Combined with the new merger_retriever, FAISS users can now perform more
granular vectorstore retrieving based on score_threshold:
```
store_1 = st[0][1].as_retriever(search_type="similarity_score_threshold", search_kwargs={"score_threshold": .8, "k": 3})
store_2 = st[1][1].as_retriever(search_type="similarity_score_threshold", search_kwargs={"score_threshold": .3, "k": 3})
store_3 = st[2][1].as_retriever(search_type="similarity_score_threshold", search_kwargs={"score_threshold": .5, "k": 3})
merged = MergerRetriever(retrievers=[store_1, store_2, store_3])
pipeline_compressor = prep_compressor()
compression_retriever = ContextualCompressionRetriever(base_compressor=pipeline_compressor, base_retriever=merged)
chain = ConversationalRetrievalChain.from_llm(OpenAI(temperature=0), chain_type="stuff", retriever=compression_retriever)
```
@dev2049 or @hwchase17 would appreciate a check if available. Thanks.
---------
Co-authored-by: Sims Juju <[email protected]>
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎
|
|
This score implementation affects similarity_search_with_relevance_scores(base.py). In similarity_search_with_score_by_vector(faiss.py), original score value is checked by 'score_threshold'. So, if call similarity_search_with_relevance_scores, score value filtering is processed twice on different criteria before and after conversion. |
This was fixed by #6570. Thanks. |
No description provided.