community: Add support for negative embeddings in PGVector

[White-RaBot]

Description: This commit introduces the capability to pass negative embeddings as a list of strings to the similarity_search_with_score_by_vector method in the PGVector vectorstore. Negative embeddings allow for refining search results by excluding certain concepts or themes.

Key changes:

• Update similarity_search_with_score_by_vector to accept a negative_embeddings parameter as a list of strings.
• Modify __query_collection to convert negative strings to embeddings and incorporate them into the search query.
• Add a filter condition in __query_collection to exclude results based on their similarity to the negative embeddings.
• Handle cases where the negative embeddings are missing or not in the expected format gracefully.

This feature enables users to have more control over their search results by specifying negative concepts to exclude. The negative embeddings are converted to vector embeddings using the provided embedding function and used to filter out unwanted results based on a configurable threshold.

Example usage:

negative_concepts = ["cats", "dogs"]
results = vectorstore.similarity_search_with_score_by_vector(
   embedding=query_embedding,
   k=10,
   negative_embeddings=negative_concepts
)

Dependencies: No additional dependencies are required for this change.

Twitter Handle: @GentJayson

[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	Jun 18, 2024 3:36am

[isahers1]

@White-RaBot Please fix errors in pgvector.py, I believe lint is yelling because there are type errors

[efriis]

closing due to no response - would also recommend doing this against the integration package in langchain-ai/langchain-postgres now instead of on the deprecated community integration!