PDF Retrieval MCP Server

A completely free Model Context Protocol (MCP) server for retrieving relevant chunks from PDF documents using hybrid search (BM25 + Vector Search).

🚀 Features

• PDF Document Processing: Automatic parsing and indexing of PDF files using Docling

• Hybrid Retrieval: Combines BM25 (keyword) and vector search (semantic) for accurate retrieval

• Free Embeddings: Uses ChromaDB's default sentence-transformers (no API costs!)

• Pure Retrieval Mode: Returns raw document chunks for agent processing (no LLM answer generation)

• Fresh Start: Clears vector database on each startup for clean indexing

• MCP Integration: Exposes retrieve_pdf_chunks tool via FastMCP for seamless agent integration

Related MCP server: PDF RAG MCP Server

📋 Prerequisites

• Python 3.11 or later

• PDF documents to index

• No API keys required! ✨

🛠️ Installation

1. Clone the Repository (if not already done)

git clone <repository-url>
cd pdf_mcpserver

2. Install Dependencies with uv

uv sync

This will automatically:

• Create a virtual environment (.venv)

• Install all dependencies from pyproject.toml

• Set up the project

3. Add PDF Documents

Create a documents directory and add your PDF files:

mkdir documents
# Copy your PDF files to the documents/ directory

That's it! No API keys or additional configuration needed.

🎯 Usage

Running the Server

uv run python main.py

Or activate the virtual environment first:

source .venv/bin/activate  # On Windows: .venv\Scripts\activate
python main.py

The server will:

1. Start immediately (lazy initialization)

2. Load and index PDFs on first query

3. Be ready to retrieve document chunks via MCP

Using the retrieve_pdf_chunks Tool

The server exposes a single MCP tool: retrieve_pdf_chunks(query: str, max_chunks: int = 5) -> str

Example Query:

retrieve_pdf_chunks("machine learning algorithms", max_chunks=3)

Example Response:

{
  "query": "machine learning algorithms",
  "chunks": [
    {
      "content": "Machine learning algorithms can be categorized into supervised, unsupervised, and reinforcement learning...",
      "document_name": "ml_guide.pdf",
      "page_number": 12,
      "metadata": {"source": "ml_guide.pdf"}
    },
    {
      "content": "Common supervised learning algorithms include linear regression, decision trees, and neural networks...",
      "document_name": "ml_guide.pdf",
      "page_number": 15,
      "metadata": {"source": "ml_guide.pdf"}
    }
  ],
  "total_chunks": 2
}

Response Structure

How Agents Use This

When an agent (like Claude) calls this tool:

1. Agent sends a search query

2. Server returns relevant document chunks

3. Agent uses chunks in its context to answer questions

Example Agent Flow:

User: "What are the main ML algorithms discussed?"
  ↓
Agent calls: retrieve_pdf_chunks("machine learning algorithms")
  ↓
Server returns: 3 relevant chunks from PDFs
  ↓
Agent reads chunks and generates answer for user

🔍 Testing with MCP Inspector

The MCP Inspector is a web-based tool for testing and debugging MCP servers interactively.

Running the Inspector

npx @modelcontextprotocol/inspector uv run python main.py

This command will:

1. Start the MCP Inspector proxy server

2. Launch your PDF Retrieval Server

3. Open a web browser with the Inspector UI

What You'll See

The Inspector provides:

• Tool Discovery: View available tools (retrieve_pdf_chunks)

• Interactive Testing: Test queries with custom parameters

• Real-time Responses: See JSON responses in real-time

• Request/Response Logs: Debug the MCP protocol communication

Example Inspector Workflow

1. Open the Inspector - Browser opens automatically at http://localhost:6274

2. Wait for Initialization - Server loads and indexes PDFs on first query (~1-2 minutes)

3. Select Tool - Click on retrieve_pdf_chunks in the tools list

4. Enter Query - Type your search query (e.g., "machine learning")

5. Set Parameters - Optionally adjust max_chunks (default: 5)

6. Execute - Click "Run" to see the results

7. View Response - Inspect the returned chunks and metadata

Inspector Tips

• First query is slow: PDF indexing happens on first query (87 seconds for typical PDFs)

• Subsequent queries are fast: Embeddings are cached in ChromaDB

• Fresh start: Server clears ChromaDB on each restart for clean indexing

• Check logs: Terminal shows detailed logging of the indexing process

🏗️ Architecture

pdf_mcpserver/
├── src/
│   ├── config.py              # Configuration management
│   ├── constants.py           # Configuration constants
│   ├── models.py              # Pydantic response models
│   ├── pdf_processor.py       # PDF loading and hybrid retrieval
│   └── retrieval_handler.py   # Document chunk retrieval
├── main.py                    # MCP server entry point
├── pyproject.toml             # Project metadata
└── .env                       # Environment configuration

Key Components

• PDFProcessor: Singleton class that loads PDFs, converts to Markdown using Docling, and builds hybrid retriever (BM25 + Vector Search)

• RetrievalHandler: Retrieves relevant chunks for queries - no LLM answer## 🔧 Configuration

Configuration is managed through environment variables. Create a .env file in the project root:

# Optional: PDF Documents Directory (defaults to ./documents)
PDF_DOCUMENTS_DIR=./documents

# Optional: ChromaDB Directory (defaults to ./chroma_db)
CHROMA_DB_DIR=./chroma_db

# Optional: Log Level (defaults to INFO)
LOG_LEVEL=INFO

Configuration Options

Note: No API keys required! ChromaDB uses free local embeddings (sentence-transformers).

🧪 Testing

Run unit tests:

uv run pytest tests/

📝 Troubleshooting

No PDF files found

Error: No PDF files found in ./documents

Solution: Add PDF files to the documents/ directory or update PDF_DOCUMENTS_DIR in .env

Import errors

Error: ModuleNotFoundError: No module named 'docling'

Solution: Ensure all dependencies are installed: uv sync

CUDA out of memory

Error: CUDA out of memory

Solution: The server is configured to use CPU-only mode. If you still see this error, check that CUDA_VISIBLE_DEVICES="" is set in src/pdf_processor.py

📚 Dependencies

• fastmcp: MCP server framework

• docling: Document processing and parsing

• chromadb: Vector database with free sentence-transformers embeddings

• langchain: RAG framework and retrievers

• loguru: Logging

No paid APIs required! All embeddings are generated locally using ChromaDB's default model (all-MiniLM-L6-v2).

🤝 Contributing

This is a Proof of Concept (PoC) implementation. For production use, consider:

• Adding caching for processed documents

• Implementing multi-agent workflow with fact verification

• Supporting additional document formats (DOCX, TXT, etc.)

• Adding authentication and rate limiting

📄 License

[Your License Here]

🙏 Acknowledgments

Based on the docchat-docling architecture.