dyj-naj / Agentic-RAG-Knowledge-Assistant

What is Agentic-RAG-Knowledge-Assistant?

This is a Python-based knowledge Q&A system that combines retrieval-augmented generation with autonomous agent capabilities. The backend runs on FastAPI while the frontend uses React with Vite. Rather than simple keyword matching, it routes every query through a 7-stage pipeline: query optimization, intent recognition, task decomposition, ReAct-style retrieval, relevance checking, answer generation, and Self-RAG quality evaluation. If the answer fails quality checks, the system loops back and retries automatically. Users get per-user knowledge bases where they can upload PDFs, Word docs, Excel files, and more -- the system handles parsing and vectorization automatically. It also pulls from four knowledge sources simultaneously: a vector database for semantic search, BM25 for keyword matching, MySQL for structured data, and web search for real-time information.

Why is it gaining traction?

The standout feature is the closed-loop error correction. Most RAG systems return whatever they retrieve; this one actually evaluates whether the answer is accurate, hallucination-free, and complete -- then retries if it fails. The ReAct retrieval agent can dynamically choose between different search strategies mid-query, which is smarter than running all retrievers in parallel and hoping for the best. Developers also get SSE streaming for real-time responses with visible processing stages, plus a three-tier conversation memory that compresses older turns to stay within context limits. The multi-source hybrid retrieval combining vector similarity with keyword matching tends to produce better results than either approach alone.

Who should use this?

Product teams building internal knowledge bases who need accurate, traceable answers will get the most value. Support teams wanting to create AI assistants grounded in their documentation should look here. Developers evaluating agentic RAG architectures for research or prototyping purposes will find a well-documented implementation to learn from. Teams needing multi-user isolation with per-user document collections will appreciate the built-in user management. However, teams needing production-ready systems should note the low star count and treat this as a reference implementation rather than a deploy-and-forget solution.

Verdict

This is a solid educational reference for agentic RAG patterns, with a complete frontend, API, and the full retrieval-evaluation loop implemented. The credibility score of 0.8999999761581421% and 11 stars indicate an early-stage project that needs community validation before production use. Start here if you want to understand how Self-RAG and ReAct retrieval work in practice, but budget time for hardening before mission-critical deployments.