@gnldev/rag

Vector store + RAG tool for AI SDK agents. Used as a tool inside runDurable, it's automatically replayable (thanks to durableTool the retrieval is journaled: on resume it comes back from the record instead of running again).

Install: pnpm add @gnldev/rag — or use it from a repo clone: pnpm install && pnpm -r build.

npm i @gnldev/rag   # peer: @gnldev/durable, ai, zod

import { InMemoryVectorStore, indexDocuments, createRagTool, llmReranker } from '@gnldev/rag';

const store = new InMemoryVectorStore();
await indexDocuments(store, embed, [{ id: 'p1', text: 'Return policy: 30 days…' }]);

const searchPolicy = createRagTool({
  store, embed, topK: 3,
  rerank: llmReranker({ model }), rerankTopK: 2,   // optional LLM reranker
});

await runDurable({ runId: 'r1', journal, model, tools: { searchPolicy }, prompt: '…' });

API

• InMemoryVectorStore · indexDocuments(store, embed, docs) · VectorStore interface (plug in your own backend)
• createRagTool({ store, embed, topK?, rerank?, rerankTopK?, description? }) → AI SDK tool
• llmReranker({ model }) → Reranker
• SemanticMemory — vector-based recall (memory integration)

How it works

Since RAG is a tool, it naturally fits into the durable agent loop: retrieval + rerank are journaled once and replayed rather than repeated. For cross-run reuse, it can be combined with @gnldev/cache.

License

Apache-2.0 — see LICENSE.