# pgvector

> Open-source vector similarity search for PostgreSQL. pgvector is a Postgres
> extension, not a service: it adds vector data types, distance operators and two
> approximate-nearest-neighbour index access methods to a database you already run.
> There is no HTTP API, no endpoint, no API key and no hosted offering. You reach it
> with SQL over an ordinary Postgres connection.

Generated by API Evangelist on 2026-08-27 from the project's own public material. This
is not published by pgvector. Method: generated. Source:
https://github.com/pgvector/pgvector

## What it is

- Extension name: `vector`. Install with `CREATE EXTENSION vector;`
- Current version: 0.8.6, released 2026-07-29. Requires PostgreSQL 13+.
- License: PostgreSQL License. Free. No paid tier, no commercial edition.
- Maintainer: Andrew Kane. Home: https://github.com/pgvector/pgvector
- The project operates no domain. pgvector.dev is a third-party redirect, not theirs.

## How to call it

There is no request to make. The canonical query is:

```sql
CREATE EXTENSION vector;
CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
```

The index is only used when `ORDER BY` is a bare distance operator in ascending order
and a `LIMIT` is present. Wrapping the operator in an expression defeats it.

## Types

- `vector` — single-precision, up to 16,000 dims, `4 * dims + 8` bytes
- `halfvec` — half-precision, up to 16,000 dims, `2 * dims + 8` bytes (0.7.0+)
- `sparsevec` — up to 16,000 non-zero elements, `8 * nnz + 16` bytes (0.7.0+)
- `bit` — Postgres' own type, given binary distance operators by pgvector (0.7.0+)

All elements must be finite. NaN and Infinity are rejected at input time.

## Distance operators

- `<->` Euclidean (L2) · `<#>` negative inner product · `<=>` cosine · `<+>` taxicab (L1)
- `<~>` Hamming and `<%>` Jaccard, for `bit`

## Indexes

- `hnsw` — graph index. Build: `m` (16), `ef_construction` (64). Query: `hnsw.ef_search` (40),
  `hnsw.iterative_scan`, `hnsw.max_scan_tuples` (20000), `hnsw.scan_mem_multiplier` (1)
- `ivfflat` — inverted list index. Build: `lists`. Query: `ivfflat.probes` (1),
  `ivfflat.iterative_scan`, `ivfflat.max_probes` (100)

Index dimension ceilings are lower than type ceilings: `vector` and `halfvec` index to
2,000 and 4,000, `bit` to 64,000, `sparsevec` to 1,000 non-zero elements under HNSW.
Set query options with `SET LOCAL` inside a transaction so they scope to one query.

## Things that fail silently

- Fewer than k results after adding an index is normal, not a bug: results are capped by
  `hnsw.ef_search` or `ivfflat.probes`. Enable iterative index scans (0.8.0+).
- NULL vectors — and zero vectors under cosine distance — are not indexed.
- A `WHERE` clause is applied after the index returns candidates, so filtering can
  return fewer rows than asked for.

## Client libraries

First-party, one per language, all under github.com/pgvector. They encode and decode the
vector types for your existing Postgres driver; they do not connect on their own.

- Python `pgvector` 0.5.0 · JavaScript/TypeScript `pgvector` 0.3.0 · Ruby `pgvector` 0.3.3
- Rust `pgvector` 0.4.2 · Go `pgvector-go` v0.4.1 · PHP `pgvector/pgvector` v0.2.2
- .NET `Pgvector` 0.3.2 · Java `com.pgvector:pgvector` 0.1.6 · Elixir `pgvector` 0.4.0
- Dart, Haskell, Nim, Julia, C++, Swift, Lua also published. 19 further repositories are
  example code only, with no package.

Full detail with versions and publication dates: packages/pgvector-packages.yml

## Reference

- Documentation and reference: https://github.com/pgvector/pgvector#readme
- Contract (the extension DDL): https://raw.githubusercontent.com/pgvector/pgvector/master/sql/vector.sql
- Changelog: https://github.com/pgvector/pgvector/blob/master/CHANGELOG.md
- Issues: https://github.com/pgvector/pgvector/issues
- Organization: https://github.com/pgvector

## What does not exist

Stated so an agent stops looking: no OpenAPI, no GraphQL, no gRPC, no AsyncAPI, no
webhooks, no MCP server, no agent card, no /.well-known/ documents, no status page, no
SLA, no rate limits, no pricing page, no SECURITY.md. Several of those are inapplicable
to a database extension. The missing SECURITY.md is not — it is a real gap.