pgvector

Model

1. TouchesModel APIs: Touches

   Model APIs, touches. pgvector stores and queries the vectors your application generates, and you can generate them in one language and query them in another. It calls no model itself.

2. Model gateway: Touches

   Model gateway, touches. pgvector runs inside the database and routes no model traffic.

Build

1. Orchestration: Touches

   Orchestration, touches. Frameworks reach pgvector through Postgres clients: the README lists pgvector-python, pgvector-node and pgvector-java among its libraries. The agent logic stays in the framework.

2. Tool use: Touches

   Tool use, touches. Any language with a Postgres client can use pgvector, so an agent's retrieval tool queries it with SQL.

Data

1. pgvectorCoreRetrieval: Core

   Retrieval, core. This is pgvector's primary role: it stores single-precision, half-precision, binary and sparse vectors in Postgres and searches them exactly by default, or with HNSW and IVFFlat indexes and iterative scans for filtered queries.

Evaluate

1. Evaluation: Touches

   Evaluation, touches. Exact nearest neighbor search, the default, gives perfect recall to compare an approximate index against. Scoring answers stays with an evaluation product.

Operate

1. Observability: Touches

   Observability, touches. pgvector runs in Postgres, so query plans and logs are Postgres's own. Application tracing stays with an observability tool.

2. SupportsCloud deployment: Supports

   Cloud deployment, supports. Neon, Supabase, Cloud SQL, Amazon RDS and Azure Database for PostgreSQL offer pgvector, each with its own release and enablement steps, and it also installs on Postgres 13 or later you run yourself.

3. Guardrails: Touches

   Guardrails, touches. The pgvector project documents no security controls of its own, so check them with the Postgres host. It names security@pgvector.org for security issues.