RAG stands for Retrieval Augmented Generation. It's a way to pull in context. This can allow for a more selective and targeted approach. Search and dump. Use

  • your eBooks
  • papers
  • conference proceedings
  • business reports (incl. PPT)
  • Excel sheets
  • ...

RAGflow is a software that makes all of that possible. Using it the right way saves a lot of time. Research time, analysis time, data prep time.

You don't need to read all of "this"... you need AI to skim through this. Grok the content. Get you the interesting stuff. Serve you.

RAGflow understands such reports and charts. It has a scanning model that is locally executed. Yes, you need a beefy server. But better let the server work, than your eyes and your brain.

From RAGflow to Claude

The problem: how to get these wonderful information into your AI tools. Like Claude Desktop or claude , or opencode . Or the newest variant.

Claude Desktop can tabularize data, and even prepare them for Excel. All you need to do is ask. And give it data.
Let RAGflow read the KPMG report the annual report and even connect the dots for you. WACC plus the debt-to-equity ratio—it's all there. Right for the analyst.
And if you need code, tools like OpenAI codex , Google gemini-cli or Anthropic claude will work for you.
claude can not only add the data to the context. It can write code to calculate model forecasts and aid you with Excel formulas.

How to get started with RAGflow and Agentic

The answer: ragflow-claude-desktop-local-mcp, named after a famous scholar from the early 1900s...

MCP (Model Context Protocol) is an immature new technology and needs further development. Do not use this in production unless you have development experience. This tool should not be used outside of research setups for now.

An MCP config for Claude Desktop can look like the following listing. The setup is based on Anaconda Python, which is used by Data Scientists (and within MS365 Python cells in Microsoft Excel).


  {
  "mcpServers": {
    "ragflow": {
      "command": "/Users/mc/anaconda3/envs/mcp-servers/bin/uv",
      "args": [
        "run",
        "--directory",
        "/Users/mc/Source/ragflow-claude-desktop-local-mcp",
        "ragflow-claude-mcp"
      ],
      "env": {
        "RAGFLOW_BASE_URL": "http://192.168.122.93:9380",
        "RAGFLOW_API_KEY": "ragflow-A1...W"
      }
    }
  }
}

Which provider and which Agents work best?

Provider Model Tier/Tool Speed Result (1-5) Note
OpenAI o3 Pro Codex + RAGflow MCP 2s 4+
OpenAI codex-mini-latest Pro Codex + RAGflow MCP FAIL 0 The model seems to have problems in general 2025-07-08
OpenRouter DeepSeek OpenRouter Codex + RAGflow MCP 2s 3
OpenRouter DeepSeek OpenRouter Claude Code CCS + RAGflow MCP 2s 3 Does not detect context window overflow
Google Gemini Free Gemini CLI + RAGflow MCP 10s 4.8
x-AI Grok 3 OpenRouter Codex + RAGflow MCP 4s 4.9
Anthropic Claude Sonnet Pro Claude Code + RAGflow MCP 4s 5 Auto-switches Claude models
Anthropic Claude Sonnet 4 GitHub Pro opencode + RAGflow MCP 4s 5 Hacky (Sonnet 4 without extra charges)

Connect the dots - with GraphRAG

GraphRAG can literally be used to connect the dots and cross-connect different reports. But sadly it's a little too expensive for me (15 Euro per 50 pages roughly).
I used these models. OpenAI embeddings work well. RAGflow uses a hybrid search. The chat model doesn't matter for API usage. I did not use the Rerank feature yet.

Instead of Rerank, I use DSPy for query optimization.

This is an optional feature that is implemented with OpenRouter or OpenAI API support. It's probably even better than reranking, and it allows simple queries to return high quality information. With the power of GPT-5, Sonnet or Opus combined.

Summary

This is a research project. It's mighty, but it's a proof-of-concept research and development code base.

It shows promising results, which speak for themselves. Personally, I find RAGflow a fantastic software, like many Chinese open-source contributions recently.