Why GraphRAG MCP

To improve information retrieval efficiency within enterprises, there is a need for an Agent capable of extracting user-relevant information from scattered documents. By building enterprise document retrieval as an MCP (Model Context Protocol), other agents within the organization can simply connect to this MCP whenever document retrieval is required. This approach centralizes document search capabilities, making it easier to integrate and scale intelligent agents across the enterprise.

How to Run

1. Fill in the required environment variables in both .env files located in the project root and in the graphrag/ directory.

2. Open your terminal and navigate to the project root directory.

3. Run the following command to start the GraphRAG MCP:

   uv run rag_client.py rag_server.py

After these steps, the GraphRAG MCP will be up and running.

Code Structure

.
├── main.py
├── rag_client.py # MCP client for graphrag
├── rag_server.py # MCP server for graphrag
├── pyproject.toml
├── .env 
├── .gitignore
├── .python-version
├── README.md
└── graphrag/
    ├── .env
    ├── settings.yaml
    ├── cache/
    │   ├── community_reporting/
    │   ├── extract_graph/
    │   ├── summarize_descriptions/
    │   └── text_embedding/
    ├── input/ 
    │   └── test.txt # knowledge base
    ├── logs/
    ├── output/
    │   ├── context.json
    │   ├── stats.json
    │   └── lancedb/
    └── prompts/
        ├── basic_search_system_prompt.txt
        ├── community_report_graph.txt
        ├── community_report_text.txt
        ├── drift_reduce_prompt.txt
        ├── drift_search_system_prompt.txt
        ├── extract_claims.txt
        ├── extract_graph.txt
        └── ...

Folder and File Descriptions

• rag_client.py: Client for interacting with the RAG server.

• rag_server.py: Server providing RAG-based tools.

• pyproject.toml: Python project configuration.

• .env, .gitignore, .python-version: Environment and versioning files.

graphrag/

• .env, settings.yaml: Environment and configuration for GraphRAG.

• cache/: Stores intermediate results for various pipeline stages.

  • community_reporting/: Caches community report data.

  • extract_graph/: Caches graph extraction results.

  • summarize_descriptions/: Caches description summaries.

  • text_embedding/: Caches text embeddings for retrieval.

• input/: Contains input data files.

• logs/: Stores log files generated during runs.

• output/: Stores output data such as context and statistics.

  • context.json: Output context data.

  • stats.json: Output statistics.

  • lancedb/: Database files for vector storage.

• prompts/: Contains prompt templates for different tasks.

  • basic_search_system_prompt.txt: Prompt for basic search.

  • community_report_graph.txt: Prompt for graph-based community reports.

  • community_report_text.txt: Prompt for text-based community reports.

  • drift_reduce_prompt.txt, drift_search_system_prompt.txt: Prompts for drift analysis.

  • extract_claims.txt, extract_graph.txt: Prompts for claim and graph extraction.

Core Technology: GraphRAG

GraphRAG is a Retrieval-Augmented Generation (RAG) framework that integrates graph-based reasoning with large language models. It enhances traditional RAG by representing entities, relationships, and claims as a knowledge graph, enabling more structured and context-aware retrieval.

Related MCP server: Graforest MCP

How GraphRAG Works

1. Data Ingestion: Raw text and tabular data are processed to extract entities, relationships, and claims, which are stored in a graph structure.

2. Prompting: Custom prompts guide the language model to generate responses grounded in the graph data.

3. Retrieval: When a query is received, relevant nodes and edges from the graph are retrieved using semantic search and graph traversal.

4. Generation: The language model synthesizes answers using both the retrieved graph context and prompt templates, ensuring responses are accurate and well-grounded.

5. Caching and Output: Intermediate and final results are cached for efficiency and stored for further analysis.

This approach allows for more explainable, reliable, and context-rich answers compared to standard RAG pipelines.