@qmilab/lodestar-adapter-vector

@qmilab/lodestar-adapter-vector

Governed Vector/RAG retrieval for the Lodestar Action Kernel. Part of Lodestar, the trust layer for AI agents.

A similarity search over a vector index is the dominant agent-memory pattern — and a poisoning surface: the chunks it returns are arbitrary stored text an attacker may have written into the index, and a crafted query can surface them on demand. This adapter makes a retrieval governed: the chunks are recorded faithfully as UNTRUSTED external_document content, so the Memory Firewall's auto-observation gate keeps a retrieved chunk from ever auto-promoting to a trusted (supported) belief.

A TS-level governance boundary, not database containment. The query reaches a real pgvector index by design; database-side privileges (a least-privileged role) are the operator's defence in depth.

Tool

What it enforces

• Retrieved chunks cannot auto-promote a belief. Paired with the VectorAwareEvidenceLinker (in @qmilab/lodestar-cognitive-core), each chunk is stamped external_document, so the firewall's Round 5 auto-observation gate keeps it unverified — even at promote-grade aggregate strength and even when several chunks corroborate. The record that a query ran may promote; the chunk text may not.
• Pinned index + namespace. The table, columns, and queryable namespaces are operator config — never agent input. The agent supplies only a query embedding and may pick only from the operator's namespace allowlist.
• Parameterized values. The embedding, the namespace, and the LIMIT are always bound; only validated, double-quoted operator identifiers are interpolated.
• Top-k cap + per-chunk content cap, bounded server-side. LIMIT bounds the number of chunks and left(content, N) bounds each chunk's text — both in the database — so a huge index, or a single poisoned/oversized row, cannot inflate an observation or balloon host memory (a trimmed chunk is flagged content_truncated). Rows with a NULL embedding are filtered, not returned as a NaN distance.
• Read-only, scoped credentials. A READ ONLY transaction with a statement_timeout; the connection password is operator config and redacted from any caught error.

Usage

import { registerVectorTools } from "@qmilab/lodestar-adapter-vector"
import {
  registerVectorRetrievalExtractor,
  VectorAwareEvidenceLinker,
} from "@qmilab/lodestar-cognitive-core"

// 1. Register the retrieval tool over an operator-pinned pgvector index.
const adapter = registerVectorTools({
  connection: { url: process.env.VECTOR_DATABASE_URL! }, // operator config; never agent input
  query: {
    table: "kb_embeddings",
    embeddingColumn: "embedding",
    idColumn: "id",
    contentColumn: "content",
    namespaceColumn: "namespace",
    namespaces: ["docs", "wiki"], // the queryable allowlist
    metric: "cosine",
    dimensions: 1536,
    maxTopK: 8,
  },
})

// 2. Register the matching extractor and wire the linker via guard.wrap()'s
//    cognitive seam so retrieved chunks stay external_document.
registerVectorRetrievalExtractor()
// guard.wrap({ cognitive: { evidenceLinkerFactory: ({ evidence, beliefs, claims }) =>
//   new VectorAwareEvidenceLinker(evidence, beliefs, claims) } })

// The consumer embeds the query (Lodestar ships no embedding model) and calls
// vector.query through the Action Kernel with { embedding, namespace, top_k }.
await adapter.close() // when done — the adapter owns its pooled connection

The agent never embeds the query, names the index, or supplies a connection string: it passes a pre-computed embedding and an allowlisted namespace, and gets back governed, gate-respecting evidence.

License

Apache-2.0