● Beta — KubeEra is in active development. Profiles are live; AI-perception measurement is rolling out.
KubeEra
P

pgvector

AI Agents
Open source · NOASSERTION

pgvector is a Postgres extension that adds a vector data type and similarity search operators (L2, inner product, cosine distance) directly into the database.

updated 2026-07-01

What it is

pgvector is a Postgres extension that adds a vector data type and similarity search operators (L2, inner product, cosine distance) directly into the database. It supports exact nearest-neighbor search and approximate search via IVFFlat and HNSW indexes. The pitch is simple: if you already run Postgres, you don't need a separate vector database to do embeddings-based retrieval — you add an extension, create an index, and query with SQL like everything else.

This puts it squarely in the RAG/AI agent stack: store embeddings alongside your existing relational data, join vector search results with normal WHERE clauses, and skip running a second data store.

Who builds it and why

Maintained as an open-source Postgres extension in C, currently at 22,019 GitHub stars with 24 contributors. That contributor count is worth noting — it's a small, concentrated group relative to the star count, which is typical for a focused C extension but means bus-factor risk is real. There's no publicly available founding date, and no CNCF maturity level (it isn't a CNCF project). It's community-driven, not backed by a foundation or a single vendor roadmap that we could verify.

Adoption has been driven by pragmatism: teams building RAG pipelines didn't want to stand up Pinecone, Weaviate, or Milvus just to store embeddings when Postgres was already sitting in their stack. pgvector rode that wave hard once LLM app development took off.

Production readiness signal

Commit activity is current (last commit 2026-06-30), so the project is actively maintained, not abandoned. 22K stars indicates broad adoption and community trust. However:

  • License is NOASSERTION. GitHub couldn't auto-detect a clear license from the repo metadata. Before you ship this in a commercial product, go verify the actual license file yourself and get legal sign-off — don't assume it's permissive just because it's popular.
  • Latest release version is not publicly available in the data we have — check the releases page directly before pinning a version in production.
  • 24 contributors is thin for a project this widely deployed. Most vector-database competitors have larger, more distributed maintainer teams.

None of this means "don't use it" — it means do your own diligence on licensing and release cadence before you commit, rather than trusting star count as a proxy for governance maturity.

Who should use this

  • Teams already running Postgres (especially managed Postgres like RDS, Cloud SQL, or Supabase) who want embeddings search without adding new infrastructure.
  • RAG applications with moderate vector volume (low millions of rows) where you want to join vector similarity with relational filters in a single query.
  • Teams that prioritize operational simplicity over squeezing out the last millisecond of ANN query latency.

Who should NOT use this

  • Teams with very large vector corpora (100M+ vectors) needing sub-10ms p99 latency at scale — dedicated vector DBs (Milvus, Qdrant) have more mature sharding and index-tuning options for this.
  • Teams needing multi-tenant vector isolation, hybrid search with built-in reranking, or advanced filtering DSLs out of the box — you'll be hand-rolling more of this in Postgres.
  • Anyone who can't tolerate the operational cost of tuning HNSW/IVFFlat build times and memory on a large existing Postgres instance — reindexing a huge vector column on a production primary is not a fun afternoon.
  • Organizations that need clear, verified licensing terms for legal/compliance sign-off before adoption — resolve the NOASSERTION issue first.

Alternatives

  • Milvus — purpose-built vector database with strong horizontal scaling and index variety, better fit for very high vector volumes.
  • Qdrant — Rust-based vector DB with a simpler operational model than Milvus and good filtering support, positioned as a pgvector alternative when you outgrow Postgres.
  • Pinecone — fully managed vector DB-as-a-service; trades control and cost for zero ops overhead.

Pricing

Fully open source. pgvector itself is free — your cost is whatever you already pay for Postgres compute/storage, plus the operational overhead of index tuning and reindexing at scale.

Frequently asked

What is pgvector?+
pgvector is a Postgres extension that adds a vector data type and similarity search operators (L2, inner product, cosine distance) directly into the database. It supports exact nearest-neighbor search and approximate search via IVFFlat and HNSW indexes.
Who builds pgvector?+
Maintained as an open-source Postgres extension in C, currently at 22,019 GitHub stars with 24 contributors. That contributor count is worth noting — it's a small, concentrated group relative to the star count, which is typical for a focused C extension but means bus-factor risk is real.
Is pgvector production ready?+
Commit activity is current (last commit 2026-06-30), so the project is actively maintained, not abandoned. 22K stars indicates broad adoption and community trust. However: License is NOASSERTION. GitHub couldn't auto-detect a clear license from the repo metadata.
Who should use pgvector?+
Teams already running Postgres (especially managed Postgres like RDS, Cloud SQL, or Supabase) who want embeddings search without adding new infrastructure. RAG applications with moderate vector volume (low millions of rows) where you want to join vector similarity with relational filters in a single query.
Who should not use pgvector?+
Teams with very large vector corpora (100M+ vectors) needing sub-10ms p99 latency at scale — dedicated vector DBs (Milvus, Qdrant) have more mature sharding and index-tuning options for this.
What are the alternatives to pgvector?+
Milvus — purpose-built vector database with strong horizontal scaling and index variety, better fit for very high vector volumes. Qdrant — Rust-based vector DB with a simpler operational model than Milvus and good filtering support, positioned as a pgvector alternative when you outgrow Postgres.
How much does pgvector cost?+
Fully open source. pgvector itself is free — your cost is whatever you already pay for Postgres compute/storage, plus the operational overhead of index tuning and reindexing at scale.