Pgvector

New
assess
First Added:July 24, 2026

pgvector adds exact and approximate vector similarity search to Postgres. We assess it when embeddings should share transactions, joins, backups, and access controls with relational application data.

Blurb

Open-source vector similarity search for Postgres

Summary

When to use:

  • Vector search belongs beside relational data and SQL filters.
  • Existing Postgres operations, backups, replication, and security should cover embeddings too.
  • One system is preferable to a separate vector service.
  • Exact search or HNSW is enough for the target corpus and latency.

When to skip:

  • Vector search dominates the workload and needs a specialized local library such as FAISS or TurboVec.
  • The corpus, query rate, or distribution model exceeds the practical limits of one Postgres deployment.
  • You need an ANN algorithm or compressed index that pgvector does not provide.

Trade-offs: pgvector keeps vectors close to business data and supports transactional updates. HNSW offers better speed-recall behavior but costs memory and build time. IVFFlat builds faster and uses less memory, but it requires training and careful list/probe tuning.

Details

Capabilities

AreaSupport
SearchExact nearest neighbor, HNSW, IVFFlat
Typesvector, halfvec, bit, sparsevec
DistancesL2, inner product, cosine, L1, Hamming, Jaccard
Database featuresACID transactions, joins, WAL, PITR, replication, SQL filters
ClientsAny language with a Postgres client

Index Choice

IndexStrengthCost
NoneExact results and simple writesFull scan cost grows with table size
HNSWStrong speed-recall trade-off; no training stepSlower builds and higher memory use
IVFFlatFaster builds and lower memory useTraining plus list and probe tuning

Create one index per distance operator class. Filtering can reduce ANN recall because filtering is applied during index scans. Iterative scans can widen the search when the first pass does not return enough matching rows.

Basic Shape 

1
2
3
4
5
6

CREATE EXTENSION vector;
CREATE TABLE items (
  id bigserial PRIMARY KEY,
  embedding vector(1536)
);
CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);

Query nearest neighbors with the operator matching the index. For cosine distance, use ORDER BY embedding <=> $1 LIMIT $2.

Comparison Points

  • Choose pgvector when relational consistency and SQL filters dominate.
  • Choose FAISS for a broad embedded ANN toolkit or GPU search.
  • Assess TurboVec when compact online ingest and local deployment dominate.

References