Skip to content

Why We Built FlowRAG

February 2026

We had a simple need: index technical documentation and query it with an AI assistant. The documentation described microservices, protocols, data flows — things that are inherently connected.

We looked at existing RAG solutions. They were all Python. They all required running a server. They all assumed you had Neo4j or Postgres running somewhere. For a TypeScript project deployed on Lambda, none of this worked.

  1. A library, not a server. Import it, use it, done. No Docker, no processes to manage.
  2. TypeScript native. Our entire stack is TypeScript. Adding Python for one feature felt wrong.
  3. Lambda-friendly. Stateless queries, fast cold starts, no persistent connections.
  4. Git-friendly storage. Commit the knowledge base to the repo. Clone and query anywhere.
  5. Knowledge graphs. Vector search finds similar text. But we needed to trace data flows: “Where does this event go? What services consume it?”

LightRAG showed us the architecture: separate KV, Vector, and Graph storage with dual retrieval (vector + graph). It’s a brilliant design. But it’s Python, it runs as a server, and it depends on external databases.

We took the architecture and rebuilt it for our world:

Aspect LightRAG FlowRAG
Language Python TypeScript
Model Server Library
Indexing Continuous Batch
Deploy Container Lambda
Storage Neo4j, Postgres SQLite, LanceDB/sqlite-vec, JSON files

FlowRAG is a fully tested TypeScript library with 100% coverage. It indexes documents, extracts entities and relations, builds a knowledge graph, and answers queries using both vector similarity and graph traversal.

You can run it locally with zero external services:

const rag = createFlowRAG({
schema,
...createLocalStorage('./data'),
});
await rag.index('./content');
const results = await rag.search('how does authentication work');

Or deploy it on AWS with S3, OpenSearch, and Bedrock:

const rag = createFlowRAG({
schema,
storage: { kv: s3Storage, vector: osVector, graph: osGraph },
embedder: new BedrockEmbedder(),
extractor: new BedrockExtractor(),
});

Same API, different backends. That’s the whole point.

FlowRAG v1 is complete. All six development phases are done. We’re now focused on documentation, examples, and community feedback before the npm release.

If you’re building RAG in TypeScript, give it a try.