GraphSense Code Graph RAG

A code analysis and retrieval system that combines graph databases, vector search, and LLMs to understand and query codebases through natural language.

It indexes JavaScript/TypeScript codebases into both a Neo4j graph database and a PostgreSQL vector database, enabling sophisticated queries about code structure, dependencies, and semantic relationships.

Features

• Multi-Modal Code Analysis: Combines graph-based structural analysis with semantic vector search.
• Natural Language Queries: Ask questions about your codebase in plain English.
• Function Discovery: Find functions based on semantic similarity and structural relationships.
• Dependency Tracking: Understand import relationships and function call hierarchies.
• AI-Powered Summaries: Automatically generates summaries for functions using LLMs.
• Real-time Analysis: Processes codebases incrementally with file watching.
• MCP Integration: Integrates with your text editor or AI agent via MCP.

Quick Start

# Navigate to your git repository
$ cd ~/path/to/repo

# Run GraphSense from within the repository
$ npx graphsense

Note: You will need environment variables declared in ~/.graphsense/.env:

Environment Variables

Required variables (must be set):

• ANTHROPIC_API_KEY - Claude API key
• PINECONE_API_KEY - Pinecone API key

Prerequisites

• Node.js 16+
• Docker
• Must be run from within a git repository (GraphSense will exit if not)
• API Keys for:
  • Anthropic (Claude) - Get API Key
  • Pinecone - Get API Key

Model Context Protocol (MCP) Integration

GraphSense provides an MCP server to integrate with AI assistants like Claude Desktop, enabling natural language queries about your codebase.

MCP Configuration

The MCP server uses stdio transport and is automatically started when you run the main application.

Starting GraphSense MCP

# Navigate to your git repository
$ cd ~/path/to/repo

# Run GraphSense from within the repository
$ npx graphsense

Claude Desktop Configuration

Add this to your Claude Desktop configuration file:

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

{
  "mcpServers": {
    "graphsense": {
      "command": "npx",
      "args": ["graphsense"],
      "env": {}
    }
  }
}

Note: Your preferred AI coding editor/agent will have similar configuration options.

Available MCP Tools

• similar_functions - Find functions based on semantic description

  • Parameters: function_description (string), topK (number, optional)
  • Returns: Array of similar functions with similarity scores

• function_callers - Find functions that call a specific function

  • Parameters: functionId (string) - The element ID of the target function
  • Returns: Array of caller functions with their IDs and names

• function_callees - Find functions called by a specific function

  • Parameters: functionId (string) - The element ID of the source function
  • Returns: Array of called functions with their IDs and names

• function_details - Get detailed information about a specific function

  • Parameters: functionId (string) - The element ID of the function
  • Returns: Function details including name, code, and summary

MCP Usage Examples

Once configured, you can use natural language queries in your AI assistant:

Example Queries

Natural Language Queries:

• "Find all functions that handle user authentication"
• "Which functions have more than 5 callers?"
• "Show me functions related to database operations"
• "What files import the authentication module?"

Structural Queries:

• Functions with high coupling (many callers/callees)
• Import dependency chains
• Orphaned functions (no callers)
• Cross-module function calls

MCP Troubleshooting

Connection Issues

1. Database Connection Errors

   # Check if PostgreSQL is running
   docker ps | grep postgres
   
   # Check if Neo4j is running
   docker ps | grep neo4j
   
   # Test database connections
   psql -h localhost -p 5432 -U postgres -d graphsense

2. Port Conflicts

   • Default PostgreSQL port: 5432
   • Default Neo4j port: 7687
   • Check docker ps output for actual ports if different

3. Environment Variables Verify required environment variables are set in ~/.graphsense/.env. If not:

   # Create a dedicated config directory
   mkdir -p ~/.graphsense
   
   # Store environment variables securely
   cat > ~/.graphsense/.env << EOF
   ANTHROPIC_API_KEY=your-key-here
   PINECONE_API_KEY=your-key-here
   EOF
   
   # Set proper permissions
   chmod 600 ~/.graphsense/.env

Common Issues

• "No functions found": Ensure your repository has been indexed first
• "Connection refused": Check if database containers are running
• "Permission denied": Verify file paths and permissions in MCP config
• "API key invalid": Confirm your Anthropic and Pinecone API keys are correct

Debugging MCP Server

# Run with debug output
DEBUG=* node build/mcp.js

# Check server logs
tail -f ~/.graphsense/logs/mcp.log

Architecture

The system uses a hybrid approach combining:

1. Neo4j Graph Database: Stores structural relationships between files and functions

   • File nodes with path properties
   • Function nodes with name and path properties
   • IMPORTS_FROM relationships between files
   • CALLS relationships between functions

2. PostgreSQL with pgvector: Stores function embeddings for semantic search

   • Function metadata and code summaries
   • Vector embeddings for similarity search
   • Hybrid dense/sparse search with reranking

3. Pinecone Vector Databases: Dual-index setup for enhanced search

   • Dense embeddings index
   • Sparse embeddings index
   • Pinecone similarity-based ranking for result optimization

Database Configuration

The system uses two databases:

• PostgreSQL with pgvector: Stores function embeddings and metadata
• Neo4j: Stores code structure and relationships

Both databases are automatically started via docker, for each repository path there will be a separate set of these 2 databases.

See ENVIRONMENT.md for complete configuration guide.

AI Models

The system uses 2 AI providers:

• Claude 3.5 Sonnet: Backup model and natural language processing
• Pinecone: Vector embeddings and similarity ranking

Vector Search

Hybrid search approach:

1. Dense vector search (semantic similarity)
2. Sparse vector search (keyword matching)
3. Result merging and deduplication
4. Pinecone similarity ranking for optimal results

Development

Project Structure

src/
├── db.ts           # Database setup and connections
├── env.ts          # Environment configuration
├── index.ts        # Main indexing logic
├── mcp.ts          # Model Context Protocol HTTP server
├── parse.ts        # Code parsing and AI processing
├── watcher.ts      # File watcher for real-time analysis
└── entrypoint.ts   # Entrypoint for the package

Building

# Compile TypeScript
npx tsc

# Watch mode for development
npx tsc --watch

License

Licensed under GPL v3.0