The important similarity_search_with_score_by_vector() function is missing from the VectorStore interface

[vemonet]

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

The VectorStore class interface should have a similarity_search_with_score_by_vector() function that can be implemented by all children classes.

No one should have to implement similarity_search_by_vector() in their VectorStore child implementation, because you can just remove the scores from the results of similarity_search_with_score_by_vector().

This is really simple to implement, and would give much more capabilities to optimize retrieval processes.

Motivation

When implementing retrievers it is really important to have access to a similarity_search_by_vector_with_score() function, because that enables to avoid having to generate multiple times vectors for the same query (when building a retriever you might want to do multiple searches with the same query and embeddings, e.g. using different filters)

But when doing a similarity search having access to the score of each match is absolutely mandatory.

Unfortunately the current interface for VectorStore only has:

• similarity_search (no score)
• similarity_search_with_score
• similarity_search_with_relevance_scores (score normalized between 0 and 1)
• similarity_search_by_vector (no score)

So the maintainers are aware that getting a score with the similarity search is important! But the function to get similarity_search_with_score_by_vector() is painfully missing 😭

And to be honest if we had to choose 1, the whole similarity_search_by_vector function is useless, because anyone can easily just remove the scores when they are provided.

It's quite elementary, a similarity search method should always provide scores (and people can ignore them or not, it's easy to remove something that is not present, but the other way is not possible). Especially that they are already computed, all we need is not throw them away!

Note that when checking at the previous version of the Qdrant vectorstore (now deprecated for one that is not). the similarity_search_with_score_by_vector() was present at some point: https://github.com/langchain-ai/langchain/blob/master/libs/partners/qdrant/langchain_qdrant/vectorstores.py#L536 and it is implemented in many langchain_community vectorstores (e.g. cassandra, couchbase, faiss...)

And I really don't see why it couldn't be part of the VectorStore interface, it would just make the system clearer while giving much more control over performance to devs that are building complex retrieval systems (generating multiple times the same vector is really bad for runtime performance, customer satisfaction, cost, energy consumption and the whole planet overall!)

Proposal (If applicable)

No one should have to implement similarity_search_by_vector in their VectorStore child implementation because you can just remove the scores from the results of

• Define similarity_search_with_score_by_vector() function in the main VectorStore class interface as NotImplemented returning a list[tuple[Document, float]]
• Define the similarity_search_with_score_by_vector() function in the main VectorStore class interface: call self.similarity_search_with_score_by_vector() and remove the score

I can implement it for you in langchain_core and the QdrantVectorStore if you want and do a PR, then it could be added incrementally to other VectorStore implementations (it's not a breaking change, just a cheap but highly empowering change). Let me know!

[dolphkin]

Do you have an estimated release date for this feature?