Pinecone vector store

Pinecone is a managed serverless vector database, selected via VECTOR_STORE_PROVIDER=pinecone.

Enable it

VECTOR_STORE_PROVIDER=pinecone
PINECONE_API_KEY=your-pinecone-api-key
PINECONE_CLOUD=aws
PINECONE_REGION=us-east-1

Install the client: pip install pinecone.

Configuration

Variable	Default	Description
VECTOR_STORE_PROVIDER	pgvector	Set to pinecone to use this store.
PINECONE_API_KEY	(none, required)	API key for the Pinecone account/index. Required.
PINECONE_CLOUD	aws	Serverless cloud placement for the created index.
PINECONE_REGION	us-east-1	Serverless region for the created index.

Notes

• Unit-tested against a mocked Pinecone client, not integration-tested against a live index. Smoke-test before production use.
• The index is created on first use with the configured dimension and the cosine metric.
• Metadata is flat (string, number, boolean, or list of strings). The authoritative payload is JSON-encoded into a _payload key that reads rebuild from, while scalar keys are also written flat so they stay filterable; a key dropped by a later update may linger as a stale flat key. Total metadata is capped at 40 KB per record.
• keyword_search is degraded to a scoped fetch plus term-overlap scoring (no BM25 index), and enumeration (list/count/delete_by_filters) runs a filtered probe query capped at 10k (the API's max top_k), so results beyond that are partial.
• get/list return valid_to = None for a live row rather than an infinity timestamp.