Production-grade Agentic RAG: 10-node LangGraph, MCP, and engineering tradeoffs

ambujsystems · reddit · 2026-08-05

Author shares deep dive into building a production-grade Agentic RAG system for Indian financial/legal documents, based on a 10-node LangGraph state graph including classifier, reject, greet, cross-questioner, web search, stock tool, retriever, generator, hallucination guard, and FastMCP server. Key decisions: Jina v3 MRL truncating vectors to 256 dims saves 75% Pinecone storage; three-tier LLM fallback (NVIDIA Nemotron, Gemini) for zero cost; pybreaker circuit breaker. Runs on free-tier services with $0/month.

Original post →

More from coding & agent

coding & agent channel →