MCP 服务器

Personal Research Assistant MCP

Enables semantic search and conversational querying across a personal research library of PDFs, DOCX, and other documents using a vector database. It provides tools for document summarization, finding related papers, and high-accuracy retrieval for AI clients like Claude Desktop.

README

📚 Personal Research Assistant MCP

A production-ready MCP (Model Context Protocol) server that enables semantic search across your personal research library. Built for AI Engineers who need fast, accurate document retrieval integrated with Claude Desktop and other AI tools.

🎯 Problem Statement

Researchers and professionals accumulate dozens of papers and documents but struggle to:

Find relevant information across multiple documents
Remember which paper contained specific insights
Connect related concepts across different sources
Spend 2+ hours daily searching for information

Traditional keyword search misses semantic connections, and reading everything is impractical.

💡 Solution

An MCP server that:

Indexes documents into a vector database using semantic embeddings
Enables Claude (or any MCP client) to query your research library conversationally
Provides sub-500ms response times with 85%+ retrieval accuracy
Includes a Streamlit dashboard for management and metrics

🏗️ Architecture

Documents (PDF/DOCX/HTML/MD)
    ↓
Document Processor → Text Chunker → Embeddings
    ↓
ChromaDB Vector Store
    ↓
├─→ MCP Server (FastMCP) → Claude Desktop
└─→ Streamlit UI → Monitoring/Testing

✨ Features

Semantic Search: Natural language queries across your entire library
Multi-Format Support: PDF, DOCX, HTML, Markdown, TXT
Fast Retrieval: <500ms query latency on 1000+ chunks
MCP Integration: Works with Claude Desktop, VS Code, and any MCP client
Metadata Extraction: Automatically extracts titles, authors, keywords
Query Logging: Track usage and performance metrics
Streamlit Dashboard: Upload, search, and visualize metrics

📊 Performance Metrics

Metric	Target	Actual
Retrieval Accuracy	85%	See METRICS.md
Query Latency	<500ms	See METRICS.md
Scale	10k+ chunks	1782+ chunks

🚀 Installation

Prerequisites

Python 3.11+
2GB RAM minimum
Git

Setup

# Clone repository
git clone https://github.com/yourusername/research-assistant-mcp.git
cd research-assistant-mcp

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Install local embeddings
pip install sentence-transformers

# Configure environment
cp .env.example .env
# Edit .env - add OPENAI_API_KEY if using OpenAI embeddings

Download Sample Data

# Download 25 AI/ML papers from arXiv
python scripts/download_sample_papers.py --count 25

Index Documents

# Index sample papers
python scripts/index_docs.py --folder ./sample_papers

# Or index your own documents
python scripts/index_docs.py --folder /path/to/your/papers --recursive

📖 Usage

Start MCP Server

python mcp_server/server.py

Configure Claude Desktop

Add to claude_desktop_config.json:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "research-assistant": {
      "command": "python",
      "args": ["/full/path/to/research-assistant-mcp/mcp_server/server.py"],
      "env": {}
    }
  }
}

Restart Claude Desktop.

Launch Streamlit UI

streamlit run ui/app.py

Opens at http://localhost:8501

🛠️ MCP Tools

`search_documents`

Semantic search across your library.

Query: "What are the challenges in RAG systems?"
Returns: Top-k results with sources, scores, and metadata

`get_document_summary`

Get quick overview of a document.

Input: Document path or title
Returns: Title, author, keywords, preview

`find_related_papers`

Find documents similar to a topic.

Query: "prompt engineering techniques"
Returns: Related papers with relevance scores

📁 Project Structure

research-assistant-mcp/
├── mcp_server/          # MCP server implementation
│   └── server.py
├── rag_pipeline/        # RAG components
│   ├── config.py
│   ├── document_processor.py
│   ├── chunker.py
│   ├── vector_store.py
│   ├── retriever.py
│   └── metadata_extractor.py
├── ui/                  # Streamlit dashboard
│   ├── app.py
│   └── pages/
├── scripts/             # CLI utilities
│   ├── index_docs.py
│   └── download_sample_papers.py
├── tests/              # Testing & benchmarks
│   ├── sample_queries.json
│   └── benchmark_performance.py
├── data/               # Data storage
│   ├── chroma_db/
│   └── query_logs/
└── docs/               # Documentation
    └── METRICS.md

🧪 Testing

# Run performance benchmarks
python tests/benchmark_performance.py

# Output: Accuracy, latency, scale metrics

🐳 Docker Deployment

# Build and run
docker-compose up -d

# Access UI at http://localhost:8501
# MCP server runs on localhost:8000

📈 Example Queries

Cross-document synthesis
"Compare different fine-tuning approaches for LLMs"
Concept exploration
"How does RLHF improve model alignment?"
Technical details
"Explain transformer attention mechanisms"
Literature review
"What are recent developments in RAG systems?"

🔧 Customization

Change Embedding Model

Edit .env:

# OpenAI (paid, best quality)
EMBEDDING_MODEL=text-embedding-3-small

# Or use local (free) by default - already configured

Adjust Chunk Size

Edit .env:

CHUNK_SIZE=1000        # Characters per chunk
CHUNK_OVERLAP=200      # Overlap between chunks

Add Document Types

Edit rag_pipeline/document_processor.py to add new file type handlers.

🐛 Troubleshooting

ChromaDB errors: Delete data/chroma_db and re-index
Import errors: Verify pip install -r requirements.txt completed
UI blank: Check browser console, try Chrome/Firefox
Slow queries: Reduce TOP_K_RESULTS in .env