nsantra

RAG MCP Server

by nsantra

AIツール

Enables Claude to perform retrieval-augmented generation using LangChain, ChromaDB, and HuggingFace models for domain-aware reasoning with PDF embedding, smart retrieval, reranking, and citation-based responses.

GitHubで見る

TL;DR:

This project implements a Retrieval-Augmented Generation (RAG) MCP Server using LangChain wrappers for ChromaDB and Hugging Face models.

Designed for seamless integration with Claude Desktop and Cursor IDE as the MCP client.

Uses a single persistent Chroma vector database with multiple collections (domains).

Automatically retrieves and ranks the most relevant context for Claude, enabling domain-aware reasoning and citation-based responses.

Project Overview
- Workflow
Features
Getting Started
- Prerequisites
- Installation
Integrations
- Claude Desktop Integration
- Cursor IDE Integration
MCP Inspector
Available Tools
Project Structure
References
License

This project implements a LangChain-powered Retrieval-Augmented Generation (RAG) pipeline hosted as a FastMCP server for integration with Claude Desktop and Cursor IDE.

It uses:

langchain_chroma.Chroma for persistent, domain-based vector stores.
langchain_huggingface.HuggingFaceEmbeddings for local or HuggingFace embedding models.
langchain_community.cross_encoders.HuggingFaceCrossEncoder for local or HuggingFace reranking models for better relevance tracking.
FastMCP — a lightweight Python interface (built on FastAPI) that exposes LangChain-based retrieval tools to any MCP client such as Claude Desktop or Cursor IDE.

Each Chroma collection represents a distinct knowledge domain or document. Claude queries are routed to the appropriate collection, which retrieves top-k results and returns relevant context and citations.

⚡Workflow:

flowchart TD
  Claude[Claude Desktop]
  MCP[MCP Server: FastMCP + LangChain]
  LangChain[LangChain Wrappers → ChromaDB + HuggingFace]
  Claude --> MCP --> LangChain --> Claude

🡅 Back to Top

PDF Embedding: Add PDFs locally or via URL directly into a chosen collection.
Smart Retrieval: Retrieve context chunks per collection or across multiple collections.
Reranking Support: Uses a HuggingFace cross-encoder reranker for better document relevance.
Document Management: List, rename, and inspect metadata for locally stored documents.
Collection Management: Create, list, and delete ChromaDB collections dynamically.
Citation Provider: Citations are generated from document metadata (e.g., page numbers, source document and path, etc.).
Self-Describing Tools: describeTools() lists all available MCP tools dynamically for introspection.

🡅 Back to Top

This MCP server exposes a set of tools that can be invoked by MCP Client to perform document and collection operations — including embedding, retrieval, metadata management, and citation generation.

For a full list of available tools, their arguments, and example usage, see the dedicated documentation:
View All Tools → TOOLS.md

🡅 Back to Top

🔧 Prerequisites

[!IMPORTANT]

Ensure Anaconda is installed, if not you can download from Anaconda and also Git (if not available, download from Github).

⚙️ Installation

Create and Activate Conda Environment

conda create -n MCP python=3.11.13 -y
conda activate MCP

Clone the Repository

git clone https://github.com/NSANTRA/RAG-MCP-Server.git
cd RAG-MCP-Server

Install Dependencies

pip install -r requirements.txt

Configure .env

# Example:

# If your system has Nvidia GPU CUDA Toolkit setup, you can set the device to cuda, otherwise set it to cpu
DEVICE = "cuda"

DOCUMENT_DIR = "C:/Users/<yourusername>/Projects/RAG-MCP-Server/Documents"
CHROMA_DB_PERSIST_DIR = "C:/Users/<yourusername>/Projects/RAG-MCP-Server/Databases"

EMBEDDING_MODEL = "C:/Users/<yourusername>/Projects/RAG-MCP-Server/Models/MiniLM"
RERANKER_MODEL = "C:/Users/<yourusername>/Projects/RAG-MCP-Server/Models/MiniLM-Reranker"

[!CAUTION] You need to mention the absolute path wherever needed.

[!TIP]

The above mentioned configuration uses local downloaded models. You can download the models using the Download Model.py python script. Change the models, if needed.

You can swap the embedding or reranker paths for any HuggingFace models.

🡅 Back to Top

[!IMPORTANT] You need to download the Claude Desktop app or Cursor IDE in order to run the MCP Server as it needs a MCP Client. You can download:

Claude Desktop

Cursor IDE

The above mentioned MCP clients automatically launches the RAG MCP Server when it’s registered in the MCP configuration file.
You do not need to run the Python script manually.

Claude Desktop Integration

🛠️ Setup Instructions

Add the following entry to your Claude MCP configuration file (typically located in your Claude Desktop settings folder).
You can find the mcp configuration file here: Settings → Developer → Edit Config to open the file.
Then, add the following JSON config:

{
  "mcpServers": {
    "RAG": {
      "command": "C:/Users/<yourusername>/anaconda3/envs/MCP/python.exe",
      "args": ["<absolute to the Main.py>"],
      "options": {
        "cwd": "absolute project root directory path"
      }
    }
  }
}

⚠️ Common Issue: If Claude fails to start the MCP server, ensure that:

The Python path points to your Conda environment’s executable.

Main.py has no syntax errors and dependencies are installed.

The cwd option matches your project root directory.

Cursor IDE Integration

🛠️ Setup Instructions

Open your project in Cursor IDE and go to File → Preferences → Cursor Setting → Tool & MCP → New MCP Server to open your MCP configuration file.
Add the following JSON entry under the "mcpServers" section (adjusting paths as needed):

{
  "mcpServers": {
    "RAG": {
      "command": "C:/Users/<yourusername>/anaconda3/envs/MCP/python.exe",
      "args": ["<absolute to the Main.py>"],
      "options": {
        "cwd": "absolute project root directory path"
      }
    }
  }
}

🡅 Back to Top

[!TIP] MCP Inspector is an official developer tool from Anthropic that lets you test, debug, and inspect any Model Context Protocol (MCP) server — including custom RAG MCP servers — without requiring Claude Desktop or Cursor IDE.

[!IMPORTANT]

To use MCP Inspector, you must have Node.js installed.

During installation, enable “Add to PATH.”

Verify your installation with node -v, npm -v and npx -v.

What It Does

Lets you call tools interactively and see raw JSON input/output.
Displays system logs, server metadata, and protocol messages.
Ideal for testing new tool definitions or debugging retrieval workflows.

Installation

You can install MCP Inspector globally using npm:

npm install -g @modelcontextprotocol/inspector

Or run directly with npx (no install needed):

npx @modelcontextprotocol/inspector

Usage

Navigate to your project root directory where Main.py is located.
Launch your MCP server via the Inspector:

npx @modelcontextprotocol/inspector python Main.py

[!TIP] (If using a Conda environment, replace python with its full path. Or, first activate the environment, and use the above command as it is.)

The Inspector will open a local web interface (usually at http://localhost:6274 ) showing:
- Input/output schemas
- Real-time logs and response traces

🡅 Back to Top

├── Main.py                 # Entry point - starts the FastMCP server
│
├── Modules/
│ ├── Config.py             # Loads env vars, sets up embeddings & reranker
│ ├── Core.py               # Document-level utilities (metadata, citation, rename)
│ ├── Database.py           # ChromaDB logic for embedding/retrieval
│ ├── Utils.py              # Helper functions (file ops, reranking)
│ └── ToolDefinition.py     # MCP tool manifests and argument schemas
│
├── .env                    # Environment configuration
├── requirements.txt        # Dependencies
└── README.md

🡅 Back to Top

LangChain RAG Workflow
LangChain Documentation — RAG
Chroma Vector Database
Chroma Docs
HuggingFace Embeddings and Cross-Encoders
Sentence Transformers
Cross-Encoder Models
Anthropic MCP & Claude Desktop
Model Context Protocol Official Site
Claude Desktop Overview

🡅 Back to Top

MIT License

Copyright (c) 2025 Neelotpal Santra

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

🡅 Back to Top