MCP-RAG

Your Personal NotebookLM for Claude Desktop

Universal RAG (Retrieval-Augmented Generation) MCP server for Claude Desktop. Index documents via CLI, search them in Claude Desktop with 0% hallucination.

---

What is MCP-RAG?

Think of it as NotebookLM for Claude Desktop:

• Index any documents: PDF, Word, PowerPoint, Excel, , TXT, MD
• Natural language search: Ask questions in Claude Desktop
• 0% Hallucination: Answers based ONLY on your documents
• 100% Local: All data stays on your computer (ChromaDB)
• Simple workflow: CLI for indexing Claude Desktop for searching

---

Architecture

  Your Documents     
  (PDF, DOCX, etc)   

           
           
    [CLI: npm run cli add]
           
           

   ChromaDB Server     Vector embeddings
   (localhost:8000)  

           
           

   MCP-RAG Server    

           
           

  Claude Desktop       You ask questions here!

Two-Part System:

1. CLI = Document management (add, delete, list)
2. Claude Desktop = Search and Q&A

---

Quick Start

1. Install

git clone https://github.com/seanshin0214/mcp-rag.git
cd mcp-rag
npm install
pip install chromadb

2. Start ChromaDB Server

Keep this running in a separate terminal:

chroma run --host localhost --port 8000

3. Add Documents (CLI)

# Add single document
npm run cli add school "path/to/regulations.pdf"

# Add multiple documents
npm run cli add research "paper1.pdf"
npm run cli add research "paper2.docx"
npm run cli add work "handbook.pptx"

Supported formats:

• Documents: PDF, DOCX, HWP, TXT, MD
• Presentations: PPTX
• Spreadsheets: XLSX, XLS

4. Configure Claude Desktop

Windows: %APPDATA%\Claude\claude_desktop_config.json

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

** Option 1: Auto-start ChromaDB (Recommended)**

{
  "mcpServers": {
    "mcp-rag": {
      "command": "node",
      "args": ["/absolute/path/to/mcp-rag/start-with-chroma.js"]
    }
  }
}

This automatically starts ChromaDB before starting MCP-RAG!

Option 2: Manual ChromaDB start

{
  "mcpServers": {
    "mcp-rag": {
      "command": "node",
      "args": ["/absolute/path/to/mcp-rag/src/index.js"]
    }
  }
}

You need to manually run chroma run --host localhost --port 8000 before starting.

Important: Use your actual installation path!

5. Restart Claude Desktop

6. Ask Questions!

In Claude Desktop:

"What does the school collection say about attendance?"

"Search the research collection for methodology"

"Show me all my collections"

---

CLI Commands

# Add document
npm run cli add <collection> <file> [-d "description"]

# List all collections
npm run cli list

# Get collection info
npm run cli info <collection>

# Search test
npm run cli search <collection> "query"

# Delete collection
npm run cli delete <collection>

Examples

# Add with description
npm run cli add school "regulations.pdf" -d "School regulations 2024"

# Add multiple files (PowerShell)
Get-ChildItem "*.docx" | ForEach-Object {
    npm run cli add MyCollection $_.FullName
}

# Check what's indexed
npm run cli list
npm run cli info school

---

MCP Tools (Claude Desktop)

When you ask questions in Claude Desktop, these tools are automatically used:

Tool	Description
search_documents	Search in specific collection or all collections
list_collections	List all available collections
get_collection_info	Get details about a collection

Note: Document addition is CLI-only, not available in Claude Desktop.

---

How It Works

Indexing (CLI)

1. Read file (PDF/DOCX/PPTX/etc)
2. Extract text
3. Split into 500-token chunks (50-token overlap)
4. Generate embeddings (ChromaDB)
5. Store in collection

Searching (Claude Desktop)

1. You ask: "What's the attendance policy?"
2. MCP-RAG searches ChromaDB
3. Returns top 5 most relevant chunks
4. Claude answers using ONLY those chunks

---

Use Cases

Students

npm run cli add math "calculus-textbook.pdf"
npm run cli add physics "lecture-notes.docx"

"Explain the concept of derivatives from my math collection"

Professionals

npm run cli add company "employee-handbook.pdf"
npm run cli add project "requirements.docx"

"What's our vacation policy?"

Researchers

npm run cli add literature "papers/*.pdf"
npm run cli add notes "research-notes.md"

"Summarize the methodology from the literature collection"

---

Features

• Multi-collection support - Organize by topic
• Semantic search - ChromaDB vector embeddings
• Source attribution - See which document/chunk
• Relevance scoring - Know how confident the match is
• Multiple file formats - PDF, DOCX, PPTX, XLSX, HWP, TXT, MD
• 100% local - No cloud, all on your machine
• 0% hallucination - Only document-based answers

---

Comparison

Feature	NotebookLM	MCP-RAG
Platform	Google Cloud	Local
AI Model	Gemini	Claude
Privacy	Cloud	100% Local
Multi-collection
CLI
Cost	Free (limited)	Free (unlimited)

---

Troubleshooting

ChromaDB Connection Error

Problem: Cannot connect to ChromaDB

Solution:

chroma run --host localhost --port 8000

Keep this terminal open!

Claude Desktop: MCP Server Not Showing

1. Check claude_desktop_config.json syntax
2. Use absolute path (not relative)
3. Restart Claude Desktop completely
4. Check ChromaDB is running

No Search Results

# Verify documents are indexed
npm run cli list
npm run cli info <collection>

# Re-index if needed
npm run cli add <collection> <file>

---

Advanced

Batch Add Files

PowerShell:

Get-ChildItem "C:\docs\*.pdf" | ForEach-Object {
    npm run cli add MyCollection $_.FullName
}

Bash:

for f in /path/to/docs/*.pdf; do
    npm run cli add MyCollection "$f"
done

Custom Chunk Size

Edit src/indexer.js:

const CHUNK_SIZE = 500;      // Tokens per chunk
const CHUNK_OVERLAP = 50;    // Overlap between chunks

Larger chunks = more context, fewer chunks
Smaller chunks = more precise, more chunks

---

Project Structure

mcp-rag/
 src/
    index.js       # MCP server
    cli.js         # CLI tool
    indexer.js     # Document processing
 chroma/            # ChromaDB data (auto-created)
 package.json
 README.md
 QUICK_START.md
 HOW_TO_USE.md

---

Requirements

• Node.js 18+
• Python 3.8+ (for ChromaDB)
• Claude Desktop (latest version)

---

Contributing

Contributions welcome! This is a universal tool that can benefit many users.

---

License

MIT License - see LICENSE

---

Credits

Built with:

• Model Context Protocol (MCP) - Anthropic
• ChromaDB - Vector database
• pdf-parse - PDF extraction
• mammoth - DOCX extraction
• officeparser - PPTX extraction
• xlsx - Excel extraction
• node-hwp - extraction

---

MCP-RAG - Your documents, Claude's intelligence, zero hallucination.