Vector Stores

PostgreSQL

As we need a backend SQL database to store conversation history and other info, using Postgres as a vector store is very attractive for us. Implementeing all this functionalities using the same technology reduces deployment overhead and complexity.

See the recipes for database configs here

pip install psycopg2-binary pgvector

# backend/config.yaml
VectorStoreConfig: &VectorStoreConfig
  source: PGVector
  source_config:
    connection_string: {{ DATABASE_URL }}

  retriever_search_type: similarity_score_threshold
  retriever_config:
    k: 20
    score_threshold: 0.5

  insertion_mode: null

top_k: maximum number of documents to fetch.

score_threshold: score below which a document is deemed irrelevant and not fetched.

insertion_mode: null | full | incremental. How document indexing and insertion in the vector store is handled.

Local Chroma

# backend/config.yaml
VectorStoreConfig: &VectorStoreConfig
  source: Chroma
  source_config:
    persist_directory: vector_database/
    collection_metadata:
      hnsw:space: cosine

  retriever_search_type: similarity_score_threshold
  retriever_config:
    k: 20
    score_threshold: 0.5

  insertion_mode: null

persist_directory: where, locally the Chroma database will be persisted.

hnsw:space: cosine: distance function used. Default is l2. Cosine is bounded [0; 1], making it easier to set a score threshold for retrival.

top_k: maximum number of documents to fetch.

score_threshold: score below which a document is deemed irrelevant and not fetched.

insertion_mode: null | full | incremental. How document indexing and insertion in the vector store is handled.