pgvector is now GA on Lambda Postgres

Every Lambda Postgres database now supports pgvector out of the box. Enable it with a single line:

CREATE EXTENSION IF NOT EXISTS vector;

That’s it. No separate vector store, no nightly sync job, no “eventually consistent” caveats. Your embeddings sit next to your data and stay transactionally consistent with it.

Why bother

We watched dozens of teams stand up a Postgres database for transactional data and then stand up a second system — Pinecone, Weaviate, a self-hosted Qdrant — for vectors. The result is a permanent two-database problem: drift between the two, double the failure modes, double the on-call surface, and a sync pipeline nobody wants to own.

If your vectors describe rows in your Postgres database, they belong in your Postgres database. The only reason it wasn’t reasonable before was performance, and pgvector 0.7 closed that gap.

What we tuned

Out of the box, pgvector defaults are conservative. We bumped a few things:

• HNSW is the default index, with m=16 and ef_construction=64.
• maintenance_work_mem is auto-sized per instance — index builds run 4–7× faster than upstream defaults.
• Parallel index builds are on by default for instances with 4+ cores.

For a 10M-row table with 1536-dim embeddings, full HNSW build takes around 4 minutes on a db-medium.

Hybrid search

The most useful pattern we’ve seen is hybrid search: full-text + vector, scored together with ts_rank and cosine similarity. We’ve added a small helper to make this less awkward.

SELECT id, title,
  ts_rank(search_tsv, query) AS lex,
  1 - (embedding <=> $2) AS sem,
  lambda.hybrid_score(lex, sem, weight => 0.6) AS score
FROM articles
WHERE search_tsv @@ query OR embedding <-> $2 < 0.4
ORDER BY score DESC
LIMIT 20;

lambda.hybrid_score is a simple weighted blend; you can also drop in your own scoring function.

Pricing

Vector storage and queries are billed at standard Postgres rates. There’s no separate vector-database SKU, no per-vector pricing, no minimums. If you’re already paying for the Postgres instance, you’re already paying for vectors.

Migration

If you’re moving from a managed vector DB, our import tool ingests Pinecone, Weaviate, and pgvector dumps directly. The CLI is lambda vectors:import. We’ll write up a longer migration guide next week.