Open-source PostgreSQL operations tool

pgvector-index-manager — safe HNSW index reconciliation

An open-source Go tool for keeping pgvector HNSW indexes aligned with a chosen configuration and filtered dataset. Review the proposed changes, build a replacement concurrently, verify it, and retire the old managed index.

Project: pgvector-index-manager
Release: Version 0.1.0
License: Apache-2.0
Language: Go

GoPostgreSQLpgvectorHNSWConcurrent DDLDocker

View source code View v0.1.0 release

About the project

pgvector-index-manager

Early release: v0.1.0. Review and test the public source against your workload before production use. No adoption or production-history claim is made.

Audit filtered-search recall

What it does

Detects HNSW drift across health, opclass, dimensions, parameters, population, and ownership metadata
Uses advisory locking and concurrent replacement builds
Verifies the replacement before publishing it under the desired name
Refuses unowned replacement by default and avoids sweeping unrelated indexes

How it works

How reconciliation works

Inspects PostgreSQL catalog state for HNSW indexes on the configured table
Compares health, vector identity, population filters, opclass, HNSW parameters, ownership, and the complete specification hash
Provides status, plan, apply, and dry-run workflows
Holds a table-scoped advisory lock through the mutation sequence
Builds replacement indexes concurrently and verifies them before publication
Records structured ownership, managed-family, and specification metadata
Protects a conflicting same-name unowned index unless replacement is explicitly authorized
Retires only superseded indexes in the same managed family, optionally after a grace period

Safety and interpretation

Review plan before apply. Catalog inspection can use read-only credentials; applying a plan requires index create, comment, rename, and drop privileges. drop_unowned is an explicit authorization, not a default sweep, and unrelated indexes must never be retired.

Search IndexLane