Pinecone ships pg_pinecone, an open-source Postgres vector extension

Pinecone released pg_pinecone on 27 October, an open-source Postgres extension, under Apache 2.0, that acts as a drop-in replacement for pgvector. It ships three engines you switch between with a single call to pinecone.set_engine(): a native engine that builds and searches its index entirely inside Postgres with no Pinecone account or network calls, a pgvector-compatible engine, and a remote engine that hands off to Pinecone’s managed service for tables that have outgrown one server.

The pitch is filtering. HNSW, the index pgvector uses, applies filters after finding neighbours, which can miss matches on a strict filter. pg_pinecone filters during the search instead. Pinecone’s own benchmark, on 1 million 768-dimension vectors, has the native engine reaching 1.000 recall at a 10% filter rate in 12.3ms, against 0.920 recall at best for HNSW’s iterative scans. At a 1% filter, it reports 0.981 recall against 0.924 for HNSW. The native index also comes in at 136MB, against 2.0GB for pgvector on halfvec and 4.1GB on full-size vectors.

CREATE EXTENSION pg_pinecone;
SELECT pinecone.create_index('docs', 'embedding', filter_column => 'tenant_id');

Why it matters: if you’re running pgvector for RAG or multi-tenant search and filtering has been costing you recall, this is a swap worth benchmarking against your own data before you reach for a separate vector database.

The caveat: the numbers above are Pinecone’s own benchmark on its own extension, not an independent test.