Add "similarity_score_threshold" option for MultiVectorRetriever class

[baptiste-pasquier]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Add the ability to filter out documents with a similarity score less than a score_threshold in the MultiVectorRetriever.

Motivation

The VectorStoreRetriever base class has a "similarity_score_threshold" option for search_type, which adds the ability to filter out any documents with a similarity score less than a score_threshold by calling the .similarity_search_with_relevance_scores() method instead of .similarity_search().

This feature is not implementend in the MultiVectorRetriever class.

Proposal (If applicable)

In the _get_relevant_documents method of MultiVectorRetriever

Replace :

langchain/libs/langchain/langchain/retrievers/multi_vector.py

Lines 63 to 68 in b20c264

	if self.search_type == SearchType.mmr:
	sub_docs = self.vectorstore.max_marginal_relevance_search(
	query, **self.search_kwargs
	)
	else:
	sub_docs = self.vectorstore.similarity_search(query, **self.search_kwargs)

With :

if self.search_type == "similarity":
    sub_docs = self.vectorstore.similarity_search(query, **self.search_kwargs)
elif self.search_type == "similarity_score_threshold":
    sub_docs_and_similarities = (
        self.vectorstore.similarity_search_with_relevance_scores(
            query, **self.search_kwargs
        )
    )
    sub_docs = [sub_doc for sub_doc, _ in sub_docs_and_similarities]
elif self.search_type == "mmr":
    sub_docs = self.vectorstore.max_marginal_relevance_search(
        query, **self.search_kwargs
    )
else:
    raise ValueError(f"search_type of {self.search_type} not allowed.")

As in the VectorStoreRetriever base class :

langchain/libs/core/langchain_core/vectorstores.py

Lines 673 to 687 in b20c264

	if self.search_type == "similarity":
	docs = self.vectorstore.similarity_search(query, **self.search_kwargs)
	elif self.search_type == "similarity_score_threshold":
	docs_and_similarities = (
	self.vectorstore.similarity_search_with_relevance_scores(
	query, **self.search_kwargs
	)
	)
	docs = [doc for doc, _ in docs_and_similarities]
	elif self.search_type == "mmr":
	docs = self.vectorstore.max_marginal_relevance_search(
	query, **self.search_kwargs
	)
	else:
	raise ValueError(f"search_type of {self.search_type} not allowed.")

[aperepel]

I just realized (after spending some quality time with the sources) that I needed this one as well. Would also need to add the type to the MultiVectorRetriever.SearchType enum

[Zizo-Vi]

pr #23539 for this feature