@galdor/memory-pgvector

A PostgreSQL + pgvector long-term memory store for galdor, implementing the core memory.Store interface. It drops in behind a Retriever like the bundled InMemoryStore, but persists vectors in a Postgres table with an HNSW cosine index, so the nearest-neighbour search runs in the database and scales to large corpora. This backend is vector-only.

Requires a Postgres server with the vector extension available, and the pg package installed (a peer dependency) when constructing from a connection string.

Usage

import { openPgVector } from "@galdor/memory-pgvector";

const store = await openPgVector({
  connString: "postgres://user:pass@localhost:5432/db",
  table: "galdor_chunks",   // optional; must match [a-z0-9_]+
  dim: 1024,
});
// …or pass your own client: openPgVector({ client: myPgPool, dim: 1024 })

await store.add([{ id: "c1", documentId: "d1", index: 0, text: "…", embedding: vec, metadata: { lang: "es" } }]);
const hits = await store.retrieve({ embedding: queryVec, k: 5, filter: { lang: "es" } });
await store.delete("d1");
await store.close();

Behavior

• open runs CREATE EXTENSION IF NOT EXISTS vector, creates the table (vector(dim) + jsonb metadata) and a doc_id btree + HNSW cosine index.
• add upserts via INSERT … ON CONFLICT (id) DO UPDATE (idempotent by id).
• retrieve ranks with the <=> cosine-distance operator (score 1 - distance, negatives dropped); filter maps to JSONB @> containment.
• delete removes every row whose document_id matches.