This repository is a local, multi-repository control center for Graph-Code RAG. Use its web dashboard to register repositories, keep each knowledge graph current with a real-time watcher, and expose all registered repositories through one MCP server.

It is built on vitali87/code-graph-rag. That upstream repository remains the reference for Graph-Code's parser, graph schema, language support, CLI workflows, and implementation details.

• A React control panel for managing multiple local repositories.
• Per-repository real-time watchers with optional initial full scans.
• A unified HTTP MCP server; every tool accepts an optional repo_path.
• Memgraph for code-structure graphs and Qdrant for semantic code embeddings.
• Dashboard access to codebase queries, semantic retrieval, embedding refreshes, process status, and live log tails.

• Python 3.12+
• uv
• Docker and Docker Compose
• Node.js and npm
• Model-provider and embedding credentials in .env

From the repository root:

uv sync --extra treesitter-full
cp .env.example .env
# Configure your model, embedding, and optional Qdrant settings in .env.
docker compose up -d
cd control_panel
chmod +x start_dev.sh
./start_dev.sh

Open:

• Dashboard: http://localhost:3003
• API and Swagger UI: http://localhost:8008/docs
• Memgraph Lab: http://localhost:3000

The start script runs the FastAPI backend and Vite frontend together. Leave that terminal open while using the dashboard. The panel starts watchers and the MCP server as managed child processes. Qdrant uses a local on-disk collection by default; configure QDRANT_HOST when using a remote Qdrant service.

1. Add an absolute repository path in the dashboard.
2. Start it with Start + full scan to build its graph and semantic index.
3. Keep Start watcher running while the repository changes.
4. Start the unified MCP server from the dashboard.
5. Configure your MCP client with http://127.0.0.1:8765/mcp.

The MCP server serves every registered repository. Pass repo_path to target a specific one; otherwise the selected default repository is used.

• get_watched_repos: registered repositories and watcher states.
• get_status: MCP configuration and service status.
• ingest_status: index freshness and pending changes.
• quick_semantic_retrieval: fast semantic code matches with source locations.
• query_codebase and query_codegraph: evidence-backed codebase answers.

For parameters and response contracts, see docs/mcp-tools.md.

• Memgraph stores repository structure and relationships, such as files, symbols, imports, and calls. It is required and defaults to localhost:7687.
• Qdrant stores semantic embeddings used by quick semantic retrieval and semantic context for codebase queries. Configure a remote instance with QDRANT_HOST, QDRANT_PORT, and QDRANT_API_KEY; without a host, Graph-Code uses a local on-disk Qdrant collection. Its corpus is deliberately focused: Python class, function, and method names with available docstrings, plus chunked Markdown; implementation bodies and other source-file contents remain available through Memgraph rather than being embedded in Qdrant.

• Memgraph must be reachable before starting a watcher or MCP service.
• The dashboard is designed for localhost use and does not provide authentication.
• Repository registrations are stored at control_panel/backend/data/repos.json.
• Use Only embeddings to rebuild Qdrant semantic embeddings without re-ingesting source files.
• Use Stop all to stop all watchers, embedding jobs, the MCP service, and the control panel backend.

For control-panel API endpoints and environment variables, see control_panel/README.md.

This fork retains the underlying Graph-Code RAG system. Consult the upstream README for:

• Supported languages and Tree-sitter setup.
• Ingestion, CLI, export, optimization, and graph-schema details.
• Provider configuration options and parser architecture.
• Contributing to the core Graph-Code RAG project.

make test-quick

See CONTRIBUTING.md for development conventions and the upstream project for the core technical documentation.