How vitali87/code-graph-rag works
Project Type & Purpose
• This is an open-source, CLI-based Retrieval-Augmented Generation (RAG) system for multi-language codebases.
• It parses source code (via Tree-sitter), ingests a knowledge graph in Memgraph, and provides an interactive natural-language interface (and batch/export/optimization commands) to query and even surgically edit code.
High-Level File Structure
• Top-level files:
• Makefile, build_binary.py, pyproject.toml, docker-compose.yaml (build & deploy)
• main.py, realtime_updater.py (entry points)
• codebase_rag/ (core logic):
• config.py, main.py (orchestrator)
• parser_loader.py + parsers/ (language-specific AST processors)
• graph_loader.py, graph_updater.py + services/ (graph ingest/query layer)
• providers/ (LLM provider abstraction)
• tools/ (file I/O, code retrieval, directory lister, shell commands)
• tests/ (comprehensive multi-lang test suite)
• grammars/ (git-submodules for Tree-sitter grammars)
• examples/, optimize/ (usage examples & optimization scripts)
Architecture Overview (from README)
• Two main components:
A. Multi-language Parser: reads a code repo, feeds AST-derived nodes/edges to Memgraph.
B. RAG CLI: takes user NL queries, orchestrates LLM(s) → generates Cypher → queries Memgraph → returns results or applies edits.
• External dependencies: Memgraph database (Docker), LLM providers (OpenAI, Google Gemini, Ollama).