RAGFlow + MCP: Turning Your Best RAG Config Into a Production Assistant

You've found your best RAG settings. Now how do you turn them into a real assistant your team uses every day?

In my previous post I covered how tools like AutoRAG and RAGBuilder can measure and find the best RAG combination (embedding, chunk size, reranker...) for your data. But those tools are measuring instruments - they tell you "this is the best config" and stop there. They are not the assistant that users talk to, upload documents to, and ask questions.

For building that assistant, the most mature open-source tool I can recommend: RAGFlow (80,000+ GitHub stars).

Document Understanding

Most RAG tools read a PDF as flat text. RAGFlow's DeepDoc engine treats the document like a human would: it preserves table structure, applies OCR to scanned pages, and understands heading hierarchy. Word, Excel, PowerPoint, scanned copies, images, web pages - it handles them all.

The Setup Logic

1. Create a separate knowledge base per department or client.
2. Upload documents.
3. Pick your embedding model and chunking template (this is where you plug in the winning settings from your measurement tools).
4. RAGFlow parses and indexes.
5. Your chat assistant is ready.

Answers come with citations - users see exactly which part of which document the answer came from, cutting hallucination risk.

MCP Support

RAGFlow can run as an MCP (Model Context Protocol) server. That means you can plug your document assistant directly into MCP-enabled tools like Claude and Cursor.

Your teammate sits in Claude and asks, "what was the penalty clause in last year's supplier contract?" - Claude searches your RAGFlow knowledge base over MCP and returns a source-cited answer from your own documents. No new interface to learn; the assistant lives inside the tools your team already uses.

Architecture, in One Line

Documents → RAGFlow (OCR + parse + chunk + index) → Knowledge bases (per department/client) → Chat UI + API + MCP → Web, Slack, or clients like Claude

Everything runs self-hosted - your data never leaves your own servers. A critical advantage for privacy and compliance (GDPR/KVKK).

The Two-Step Recipe

• Use measurement tools (AutoRAG, RAGBuilder) to find the best RAG settings for your data.
• Build your knowledge base in RAGFlow with those settings, and connect your assistant to your team's tools via MCP.

A document assistant built on measurement instead of guesswork, with citations, running on your own servers - fully possible today with open-source tools alone.