pgvector support for Python, providing integrations for Django, SQLAlchemy, SQLModel, and several PostgreSQL drivers and async clients. It helps you store vector embeddings in Postgres and run operations like nearest-neighbor search, distance calculations, and vector indexes for approximate retrieval.
Project status
- Actively maintained: The source received its most recent updates on 2026-07-06, and there have been no long periods of inactivity since then (today is 2026-07-19).
- Update cadence: Updates appear on a roughly 6-8 month rhythm (0.4.1 on 2025-04-26, 0.4.2 on 2025-12-05, 0.5.0 on 2026-07-06), with the latest update very recent.
AI summary generated
Recent updates
v0.5.0
v0.5.0 introduces experimental type hint support and significantly changes how vector-related types are imported and represented across frameworks. The published release notes for this version are missing, but the diff shows multiple breaking API and behavior changes, including type return-shape changes, import path changes, and dropped environment support.
Featuresv0.4.2
Release v0.4.2 (no publisher release notes provided) updates the project for broader environment compatibility and adds/extends functionality around binary quantization. The code diff shows changes to Django integration, SQLAlchemy BIT handling for asyncpg, documentation/tests around binary quantization, and a notable change in SQLAlchemy index configuration (removal of explicit HNSW build parameters).
BreakingFeaturesv0.4.1
The v0.4.1 release was published on 2025-04-26, but no release notes were provided by the publisher. As a result, there is no documented information about new features, fixes, security updates, or potential breaking changes for this version.
v0.4.0
v0.4.0 introduces a new pg8000 integration, a top-level pgvector package API (Vector, HalfVector, Bit, SparseVector), and additional Bit/array handling improvements. It also drops support for Python < 3.9 and changes default behavior for psycopg2 type registration, which can alter how vectors are adapted globally vs per-connection.
BreakingFeaturesv0.3.6
v0.3.6 introduces support for psycopg2 casting of vector array types. The code change adds an `arrays` option to `pgvector.psycopg2.register_vector`, enabling registration of `vector[]`, `halfvec[]`, and `sparsevec[]` adapters when requested, and updates SQLAlchemy README examples for `avg` usage.
Featuresv0.3.5
Release v0.3.5 introduces SQLAlchemy support for an `avg` function (and exports `sum`) intended to provide correct SQLAlchemy typing/casting. It also adds a `globally` option to the psycopg2 registration path and updates docs/tests to use the new SQLAlchemy function and psycopg2 scoping behavior.
BreakingFeaturesv0.3.4
Release v0.3.4 includes minimal code changes. No release notes were provided by the publisher, so the only described changes are inferred from the repository diff itself.
Featuresv0.3.3
This release (v0.3.3) appears to focus on improving Psycopg2 cursor factory handling, along with significant reorganization of the repository examples. The published release notes were not provided in the release metadata, so the only explicit documentation of intent is found in the repo changelog and README updates.
v0.3.2
v0.3.2 includes a targeted runtime fix for the asyncpg integration when working with older pgvector versions, plus some documentation and example updates (including a new Cohere binary embeddings example). The provided release notes block contains no content, so the actual behavioral change is best inferred from the code diff.
v0.3.1
Release v0.3.1 was published on 2024-07-11, but no release notes or changelog details were provided by the publisher. As a result, this release cannot be assessed from the notes alone for new features, bug fixes, breaking changes, security, performance, or migration steps.