Skip to content

v0.2.0: pgContext - Composable Retrieval for PostgreSQL

Latest

Choose a tag to compare

@evokoa-admin evokoa-admin released this 23 Jul 12:26
v0.2.0

pgContext 0.2.0 expands PostgreSQL-native vector retrieval while keeping ordinary PostgreSQL tables authoritative. Transactions, MVCC, ACL/RLS, backups, and recovery continue to be owned by PostgreSQL; HNSW indexes remain rebuildable acceleration artifacts.

Highlights

  • Exact dense-vector search with L2, inner-product, cosine, and L1 metrics
  • Page-native HNSW indexes with inserts, VACUUM, REINDEX, restart, and WAL-aware lifecycle support
  • Metadata-filtered exact and approximate search
  • Collections, scrolling, counts, facets, grouping, recommendation, and discovery
  • Hybrid dense and PostgreSQL full-text retrieval
  • Named dense and sparse retrieval paths
  • Experimental halfvec, sparsevec, and bitvec types and HNSW operator classes
  • Quantization, late-interaction, and composite-query building blocks
  • Expanded diagnostics, recall checks, lifecycle status, and bounded telemetry
  • PostgreSQL 17 and 18 images for linux/amd64 and linux/arm64

PostgreSQL 17 remains the primary performance and deep-lifecycle qualification target.

Upgrading from 0.1.0

A standalone 0.1.0 → 0.2.0 extension upgrade is supported:

ALTER EXTENSION pgcontext UPDATE TO '0.2.0';

The upgrade requires a PostgreSQL superuser. pgContext-owned vector types now live in the pgcontext schema, so applications should use qualified names such as pgcontext.vector(1536) or add pgcontext to search_path.

Because the HNSW on-disk format remains experimental, rebuild existing HNSW indexes after upgrading.

For pgvector coexistence, use the optional pgcontext_pgvector companion. Its certified profile is PostgreSQL 17 with pgvector 0.8.x.

Current boundaries

This release is intended for prototypes, evaluation, and controlled pilots.

  • HNSW, filtered ANN, non-dense vectors, quantized serving, and late-interaction paths remain experimental.
  • IVFFlat is not implemented.
  • pgContext is not a drop-in replacement for pgvector.
  • Broad production and long-duration workload certification remains ongoing.

See the full release notes, installation guide, and known issues for details.