semantic-sf-docs-rag-mcp

A privacy-first, fully local Model Context Protocol server for semantic search over Salesforce documentation. Uses sqlite-vec + all-MiniLM-L6-v2 embeddings — $0 API costs, no data leaves your machine.

Ships with a pre-populated vector database and bundled ONNX model — zero configuration required.

Quick Start — npx (Recommended)

No install needed. Add to your MCP client config and it runs automatically.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (Mac) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "sf-docs": {
      "command": "npx",
      "args": ["-y", "semantic-sf-docs-rag-mcp@latest"]
    }
  }
}

VS Code Agentforce Vibes

Add to .vscode/mcp.json in your workspace (or User Settings → MCP):

{
  "servers": {
    "sf-docs": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "semantic-sf-docs-rag-mcp@latest"]
    }
  }
}

Windows users: VS Code may not inherit PATH. Use the full path to Node if needed:

"command": "C:\\Program Files\\nodejs\\node.exe",
"args": ["C:\\Users\\YOU\\.npm-cache\\semantic-sf-docs-rag-mcp\\dist\\index.js"]

Restart Claude / reload VS Code. The semantic_search_docs tool is now available.

Available MCP Tools

| Tool | Description | |---|---| | semantic_search_docs | Semantic similarity search across indexed Salesforce docs | | list_doc_sources | List all indexed documents with chunk counts | | get_source_content | Retrieve all chunks from a specific document | | add_pdf_to_index | Embed a local PDF into the database at runtime |

Architecture

| Component | Technology | |---|---| | Embedding Model | @xenova/transformers · all-MiniLM-L6-v2 (384-dim, local ONNX, ~22MB) | | Vector Database | better-sqlite3 + sqlite-vec — ships as data/rag.sqlite | | MCP Protocol | @modelcontextprotocol/sdk over STDIO |

Extending the Database with Your Own PDFs

The bundled database covers a curated set of Salesforce documentation. To add your own content:

Option 1 — Drop PDFs and run the ingest script

# Clone the repo
git clone https://github.com/TMTrevisan/semantic-sf-docs-rag-mcp.git
cd semantic-sf-docs-rag-mcp
npm install

# Add your PDF files
cp ~/Downloads/MyCustomDoc.pdf ./pdfs/

# Run the ingest (embeds all PDFs in ./pdfs/)
npm run ingest-pdfs

The script will:

• Parse text from each PDF
• Split into overlapping 800-character chunks
• Generate all-MiniLM-L6-v2 embeddings locally
• Store everything in data/rag.sqlite

Point your MCP config at the local dist/index.js to use your custom database:

{
  "mcpServers": {
    "sf-docs": {
      "command": "node",
      "args": ["/path/to/semantic-sf-docs-rag-mcp/dist/index.js"]
    }
  }
}

The server always loads data/rag.sqlite relative to its own install location. Override with the SF_DOCS_DB_PATH environment variable if needed.

Option 2 — Add a PDF at runtime via MCP tool

Ask your AI assistant to index a specific PDF without restarting the server:

Use the add_pdf_to_index tool to embed /Users/me/Downloads/AgentforceDocs.pdf

Environment Variables

| Variable | Default | Description | |---|---|---| | SF_DOCS_DB_PATH | data/rag.sqlite next to the package | Override the database path |