Built this: Enterprise Agentic RAG Orchestrator (Multi-Agent-RAG)

On 20 April 2026, in the middle of my LangChain phase, I started Multi-Agent-RAG. The repo now has 46 commits, with a Python backend and a TypeScript frontend.

Architecture

From the README:

• A FastAPI gateway checks an API key, applies IP-based rate limiting (Redis), runs a prompt-injection guard, and looks up a semantic cache (Qdrant ANN, cosine ≥ 0.96) and chat history (Redis, 1-hour TTL).
• A Supervisor LLM classifies each message into one of four routes:
  • Retriever agent: Corrective RAG in LangGraph. An LLM grader checks the relevance of retrieved documents, and the agent rewrites the query and retrieves again, at most two times.
  • DB agent: natural language to SQL, read-only, with a forbidden table.
  • Tool agent: email actions under a Draft → Confirm → Execute human-in-the-loop flow, with a domain allow-list and hourly caps.
  • Chat agent: plain conversation.
• Retrieval is hybrid: 384-dimensional dense vectors in Qdrant plus BM25, reranked by a CrossEncoder (ms-marco-MiniLM-L-6-v2). Small 200-token child chunks are retrieved and expanded to 800-token parent context.
• A Validation agent re-grounds every answer before it’s returned.

Reported quality

The README lists RAGAS results of 0.81 faithfulness, 1.00 context precision, 0.90 context recall and 0.75 answer relevancy on its evaluation datasets.

Where it connects

It pulls together my RAG and LangGraph notes, and it’s the multi-agent system I come back to in the multi-agent systems post.