Facts read November 7, 2026
pgvector (in Postgres) vs turbopuffer
pgvector (in Postgres) is Cuely's default vector search pick when your description doesn't favor either one: no extra vendor: vectors live in the Postgres you already have, so tenant filters are a plain WHERE clause. Pick turbopuffer if you care most about high volume.
Side by side
| Fact | pgvector (in Postgres) | turbopuffer |
|---|---|---|
| What it is | An open-source Postgres extension that stores and searches embeddings next to your other data. | Vector and full-text search on object storage, priced by usage. |
| Free plan | Free: on Neon it is "available on every Neon plan with no add-on or paid tier required"; you pay only for database storage and compute. | None |
| Paid from | Not stated | Launch: $16/month minimum usage; storage up to $0.33 per GB. |
| At 1,000 users | $0/month | $16/month |
| Open source | Yes | No |
| Self-hostable | Yes | No |
| Best for | Up to a few million chunks; Filtering vector search by tenant with SQL; Fewest moving parts | Very large or many-tenant indexes; Low storage cost at scale |
| Watch out | Its HNSW index supports vectors up to 2,000 dimensions (halfvec up to 4,000); pick an embedding size that fits. Vectors count toward your database storage (Neon Free: 1 GB per project). | No free plan: the cheapest plan has a $16/month minimum. |
Where they differ
The published scores, from −3 to +3, that decide which one Cuely picks for each need.
| Need | pgvector (in Postgres) | turbopuffer |
|---|---|---|
| keep costs as low as possible | +3 No separate bill; it uses the database you already pay for. | −2 No free plan; $16/month minimum. |
| high volume | −1 Past tens of millions of vectors, a dedicated vector store is easier to run. | +3 Object-storage based, so very large indexes stay cheap. |
| fewer vendors | +3 Vectors stay in your existing Postgres; no new vendor or sync job. | 0 No particular strength here. |
| prefers open source or self-hosting | +3 Open-source extension that runs on any Postgres. | 0 No particular strength here. |
| a database per customer | 0 No particular strength here. | +1 Many small namespaces (one per customer) are a common pattern. |
Sources: pgvector (in Postgres) pricing (read November 7, 2026), turbopuffer pricing (read November 7, 2026). No vendor pays for placement.