RAG Document Server

# MCP Server Setup for Claude ## Overview This guide shows how to configure Claude Desktop or Claude Code to use the RAG MCP server for querying Dagster documentation. ## Configuration ### For Claude Desktop Add this configuration to your Claude Desktop config file: **macOS:** `~/Library/Application Support/Claude/claude_desktop_config.json` **Windows:** `%APPDATA%\Claude\claude_desktop_config.json` **Linux:** `~/.config/Claude/claude_desktop_config.json` ```json { "mcpServers": { "rag-dagster-docs": { "command": "uv", "args": [ "--directory", "/home/ubuntu/work/rag", "run", "python", "-m", "mcp_server.server" ] } } } ``` **Note:** The API key is read automatically from `/home/ubuntu/work/rag/.env` - no need to put it in the config! ### For Claude Code (Current Environment) Since you're already in Claude Code, you can add the MCP server configuration: 1. **Option A: Using Claude Code Settings** - Run the command: `/mcp add` - Or manually edit the MCP configuration in Claude Code settings 2. **Option B: Direct MCP Configuration** Add to Claude Code's MCP servers configuration: ```json { "name": "rag-dagster-docs", "command": "uv", "args": [ "--directory", "/home/ubuntu/work/rag", "run", "python", "-m", "mcp_server.server" ] } ``` **Security Note:** The MCP server automatically loads the Google API key from `/home/ubuntu/work/rag/.env`. You don't need to (and shouldn't) include API keys in the MCP configuration JSON. ## Available MCP Tools Once configured, Claude will have access to these tools: ### 1. `query_rag` - Query the documentation ``` Query the RAG system with a question about Dagster. ``` **Parameters:** - `question` (required): Your question about Dagster - `top_k` (optional): Number of relevant chunks (default: 5) - `tags` (optional): Filter by tags like `["dagster", "docs"]` - `section_path` (optional): Filter by section like `"getting-started > Concepts"` **Example usage in Claude:** ``` Use the query_rag tool to ask: "How do I define assets in Dagster?" with tags: ["dagster"] ``` ### 2. `add_document` - Add new documentation ``` Add a markdown or text document to the RAG system. ``` ### 3. `list_documents` - List all documents ``` See all documents in the system, optionally filtered by tags. ``` ### 4. `delete_document` - Remove a document ``` Delete a document by its ID. ``` ### 5. `get_tags` - List available tags ``` Get all unique tags across all documents. ``` ### 6. `get_document_structure` - View table of contents ``` Get the hierarchical section structure of a specific document. ``` ### 7. `get_rag_stats` - System statistics ``` Get statistics about the RAG system (documents, chunks, etc.). ``` ## Querying Only Dagster Docs To ensure Claude only searches Dagster documentation, use the `tags` parameter: ```json { "question": "How do I create a schedule in Dagster?", "tags": ["dagster", "docs"] } ``` All Dagster docs were ingested with tags `["dagster", "docs"]`, so filtering by these tags ensures you only get Dagster-specific results. ## Testing the Setup After configuration, restart Claude Desktop/Code and try: 1. Check available tools: Claude should show the MCP tools in its tool list 2. Test a query: Ask Claude to "Use query_rag to find information about Dagster assets with tags: dagster, docs" 3. Verify results include hierarchical paths like `getting-started > Concepts > Asset` ## Environment Variables The MCP server reads configuration from `/home/ubuntu/work/rag/.env`. This file should contain: ```bash GOOGLE_API_KEY=your-actual-api-key ``` **Current status:** ✅ Already configured and working **Security best practice:** - ✅ API keys in `.env` file (gitignored) - ✅ MCP config has NO hardcoded secrets - ✅ Settings automatically load from `.env` via Pydantic To verify your `.env` is set up correctly: ```bash cd /home/ubuntu/work/rag grep GOOGLE_API_KEY .env ``` ## Troubleshooting ### MCP Server Won't Start - Check that the Google API key is set - Verify the path `/home/ubuntu/work/rag` is correct - Check logs in Claude Desktop/Code ### No Results Returned - Verify documents were ingested: Use `get_rag_stats` tool - Check tags match: Use `get_tags` tool to see available tags - Try without tag filtering first ### Wrong Documentation - Always include `"tags": ["dagster"]` in your queries - Use `section_path` to narrow down to specific sections like `"getting-started"`