Vectorstore relevance fn fix

[dangoldbj]

Description:

• The changes fixes the normalization in the relevance score functions.
• Some vector databases use cosine to indicate cosine similarity while some use it to indicate cosine distance.
  • SemaDB uses cosine metric to indicate cosine distance. https://rapidapi.com/semafind-semadb/api/semadb/details
  • Elasticsearch uses cosine metric to indicate cosine similarity. https://www.elastic.co/guide/en/elasticsearch/reference/current/dense-vector.html#dense-vector-params
  • Neo4j uses cosine metric to indicate cosine similarity. https://neo4j.com/docs/cypher-manual/current/indexes-for-vector-search/#indexes-vector-create
  • Momento uses cosine_similarity metric to indicate cosine similarity. https://github.com/momentohq/client-sdk-python/blob/main/src/momento/requests/vector_index/similarity_metric.py
  • pgvector uses cosine distance. https://github.com/pgvector/pgvector#pgvector
  • Pinecone uses cosine metric to indicate cosine similarity. https://docs.pinecone.io/docs/indexes#distance-metrics
  • FAISS does not directly support cosine similarity nor cosine distance. https://github.com/facebookresearch/faiss/blob/main/faiss/MetricType.h#L44
  • TimescaleVector is built on top of pgvector, hence uses cosine distance.
  • Chroma uses cosine metric to indicate cosine distance. https://docs.trychroma.com/usage-guide#changing-the-distance-function
  • Redis uses cosine metric to indicate cosine distance. https://redis.io/docs/interact/search-and-query/advanced-concepts/vectors/

Issue:
#11587
#13437
#7384
#11386
#12697

Twitter handle: dangoldbj

@baskaryan, @eyurtsev, @hwchase17

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 26, 2023 3:53pm

[younes-io]

@dangoldbj : is OpenSearch included in this?

[dangoldbj]

@dangoldbj : is OpenSearch included in this?

@younes-io I think OpenSearch vector store is correctly implemented.
https://opensearch.org/docs/latest/search-plugins/knn/approximate-knn/#spaces does use cosineSimilarity to indicate cosine distance. However in the implementation, https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/vectorstores/opensearch_vector_search.py#L270C5-L270C5 the score is converted back to cosine similarity

[younes-io]

you're right..
btw, i don't see in OpenSearch vector store implementation a way to do hybrid search (RRF).. I want to do keyword search and use the similarity search and combine them in an Ensemble retriever.. Ofc, I wish there was a way to call search pipelines using OpenSearch vector implementation so that OpenSearch takes care of all of the hybrid search directly since it supports since 2.10 but LangChain doesn't seem to support it for now, right ? Have you done sth like this using OpenSeachVectorSearch ?
thank you

[hwchase17]

this has a lot of merge conflicts so i am going to close. we will revisitng our vectorstore abstractions in ~1 month and will reference this