AI agents · OpenClaw · self-hosting · automation

Quick Answer

Best Postgres Vector Search Extensions in 2026, Compared

Published:

The short answer

Start with pgvector; add pgvectorscale when your index outgrows RAM or filtered queries get slow; pick ParadeDB if you need keyword and vector hybrid search in Postgres; consider VectorChord only for 100M+ vectors where storage cost dominates. All four run inside Postgres, so your embeddings stay in the same transaction as your data. Versions and licences were read from each project’s GitHub repository on November 22, 2026.

The comparison

pgvectorpgvectorscaleParadeDB (pg_search)VectorChord
Latest release0.8.7 (Oct 1, 2026); 0.9.0 in development0.9.1 (Sep 4, 2026)v0.27.0 (Nov 19, 2026)1.1.1 (Feb 28, 2026)
LicencePostgreSQLPostgreSQLAGPL-3.0 (Enterprise commercial)AGPLv3 or Elastic License v2
IndexHNSW, IVFFlatStreamingDiskANNTantivy-based index (vector search in beta since 0.25)IVF + RaBitQ quantization
StrengthEverywhere, simpleLarge indexes on disk, filtered searchBM25 + vector hybrid, facetsCost per vector at 100M+ scale
FilteringIterative index scans, partial indexesLabel-based filtered DiskANNSQL filters inside the search indexPrefilter support
GitHub activityCommits Nov 22, 2026Last push Oct 14, 2026Commits Nov 22, 2026Last push Aug 6, 2026
Pick it forDefault RAG store10M–100M vectorsSearch products, hybrid RAGBillion-scale experiments

Picks in detail

pgvector — the default. Store vector, halfvec, bit and sparsevec columns, build CREATE INDEX ... USING hnsw (embedding vector_cosine_ops), and query with ORDER BY embedding <=> $1 LIMIT 10. Indexed limits: vector up to 2,000 dimensions, halfvec 4,000, bit 64,000, sparsevec 1,000 non-zero elements — use halfvec for 3,072-dimension embedding models. The 0.9.0 changelog (unreleased) adds asynchronous I/O for indexes and drops Postgres 13 and 14. It is preinstalled on Postgres.app and most hosted providers.

pgvectorscale — when HNSW stops fitting in memory. Its StreamingDiskANN index keeps most of the graph on SSD, Statistical Binary Quantization compresses vectors, and label-based filtering applies tags during the graph search instead of after it. It is written in Rust (pgrx) and available on Timescale Cloud; self-host by building from source or using the TimescaleDB Docker image.

ParadeDB — hybrid search in one query. ParadeDB adds BM25 full-text scoring, facets and aggregations to Postgres, and indexes pgvector columns in the same index for hybrid ranking. Vector search is beta; ParadeDB itself recommends 0.27.0 or later for vector workloads. The community edition is AGPL-3.0, which matters if you ship a modified version as a service.

VectorChord — cheap at huge scale. The project claims 100 million 768-dimension vectors on one AWS i4i.xlarge ($247/month) and indexing 100 million vectors in 20 minutes (vendor figures). Development has slowed — the repository moved to supervc-stack and its last push was August 6, 2026 — so test the upgrade path before committing.

How to choose in four questions

  1. How many vectors? Under ~10M: pgvector HNSW. 10M–100M: pgvectorscale. Beyond that, benchmark VectorChord or a dedicated vector database.
  2. Do users type keywords? If product names, codes or exact phrases matter, you need BM25 too: ParadeDB, or pgvector plus Postgres full-text search with reciprocal rank fusion.
  3. Heavy metadata filters? Turn on pgvector iterative scans first; if recall still drops, pgvectorscale’s label filtering.
  4. Licence constraints? pgvector and pgvectorscale are permissive; ParadeDB and VectorChord need a licence review for SaaS use.

Embedding choice matters as much as the index — see best embedding models. For managed Postgres hosts, see Neon vs Supabase vs Prisma Postgres for AI apps; for dedicated engines, best vector databases.

Last verified: November 22, 2026.

Sources