pg_pinecone: open source search for your Postgres data

AI that knows your business starts with your own data, and for many teams that data lives in Postgres.

Today we're open-sourcing pg_pinecone, a Postgres extension (Apache 2.0) built on the search engine Pinecone runs at scale. It works with the vector columns you already have. Your queries stay plain SQL.

Instead of HNSW's graph, its index stores a compact summary of each vector. That makes it a fraction of the size, faster to build, and more accurate when a filter narrows the search to one user or tenant.

In our tests on 1 million 768-dimension vectors, with pg_pinecone on full-size vectors and pgvector's HNSW on halfvec (its half-precision format):

• Index size: 136 MB vs. 2.0 GB
• Search at 0.95 recall: 1.3 ms vs. 1.7 ms

It runs inside Postgres 17 or 18 with no account or API key. If a table outgrows one server, a Pinecone API key and one SQL call move its search to Pinecone. Your queries don't change.

Install pg_pinecone on GitHub →

Benchmarks and design notes: Read the announcement →

Pinecone Systems, Inc, 127 W. 26th St., Floor 6, New York, New York 10001, United States