Turn your monorepo into a queryable knowledge graph

What it does

Code-Graph-RAG uses Tree-sitter to parse codebases in nine languages into an abstract syntax tree, then ingests functions, classes, call graphs, and imports into Memgraph. The result is a queryable knowledge graph of your repo’s structure. An interactive CLI translates natural language questions into Cypher via an LLM, retrieves actual source snippets, and can perform AST-targeted code edits with visual diffs.

The interesting bit

Instead of flat vector search, it stores real semantic relationships in a graph database, so you can trace call chains across a monorepo. The system also lets you mix LLM providers—perhaps Google Gemini for orchestration and a local Ollama model for Cypher generation—keeping sensitive code on-premise while using cloud smarts for planning. It even handles surgical edits by targeting specific AST nodes rather than dumping raw text replacements.

Key highlights

• Supports nine languages via Tree-sitter (C, C++, Java, JavaScript, Lua, PHP, Python, Rust, TypeScript); Go and Scala are in development.
• Stores structure in Memgraph, enabling relationship-aware queries—call graphs, class hierarchies, imports—that flat search misses.
• Configurable LLM backends (OpenAI, Google Gemini, local Ollama) with separate models allowed for orchestration and Cypher generation.
• Surgical code editing uses AST-based function targeting and visual diffs, plus shell command execution for test runs or CLI tasks.
• Workspaces bundle multiple repos into one queryable graph; indexed data persists in named Docker volumes across restarts.

Caveats

• The project recently migrated to Codeberg after a GitHub account suspension, leaving several CI badges and integrations broken or outdated.
• Go and Scala support is still listed as “In Development” in the language matrix.
• It requires a local Docker stack (Memgraph and Qdrant) plus system dependencies like cmake and ripgrep, making setup heavier than a typical Python CLI.

Verdict

Worth exploring if you manage large, polyglot monorepos and need relationship-aware answers that simple text search cannot provide. Skip it if you want a lightweight, zero-dependency tool or if running a local graph database feels like overkill for your project.