PGVECTOR

pgvector PostgreSQL extension vector similarity native.

Définition

pgvector features : (1) Vector type : CREATE TABLE items (embedding vector(1536)); supports up to 16000 dimensions (HNSW indexable limit 2000), supports half-precision (halfvec 16-bit), binary (bit), sparsevec sparse. (2) Distance operators : <-> (L2 distance), <#> (negative inner product), <=> (cosine distance), <+> (L1 Manhattan), <~> (Hamming for binary), <%> (Jaccard binary). (3) Index types : IVFFLAT (Inverted File flat, faster build slower query), HNSW (Hierarchical Navigable Small World, slower build faster query, recommended). (4) Index parameters : HNSW (m, ef_construction, ef), IVFFLAT (lists). (5) Filtering : combined with SQL WHERE clauses standard, JOIN with regular relational tables (vendors, users, products etc.). (6) Quantization : Scalar Quantization SQ via half-precision halfvec 16-bit 2x compression, Binary Quantization via bit type 32x compression. (7) Performance : pgvector 0.7+ ~50% faster HNSW vs 0.5, comparable Pinecone within 2x throughput. Cloud-managed PostgreSQL support : AWS RDS PostgreSQL (since 15.2 native), Azure Database for PostgreSQL (Flexible Server), GCP Cloud SQL PostgreSQL, Aiven, Neon, Supabase, Crunchy Data Postgres, Timescale.

Origine

pgvector publie 2021 par Andrew Kane (Postgres community contributor) ; pgvector 0.5 ajoute HNSW septembre 2023 (major milestone) ; pgvector 0.7 juillet 2024 quantization halfvec + bit ; ~most deployed vector DB 2024 due cloud-managed Postgres ubiquity.

Exemple en contexte

SaaS startup B2B uses AWS RDS PostgreSQL 16 + pgvector 0.7 : embed customer support tickets via OpenAI text-embedding-3-small (1536 dim), HNSW index, SQL query 'SELECT * FROM tickets ORDER BY embedding <=> $1 LIMIT 10 WHERE customer_id = $2' returns top-10 semantically similar tickets joined customer context, no separate vector DB to manage.

Termes liés

• Redis Stack — alternative in-memory.