Code-Graph-RAG

A graph-based RAG system that parses multi-language codebases with Tree-sitter, builds knowledge graphs in Memgraph, and enables natural language querying, editing, and optimisation.

Install

pip install code-graph-rag

With all Tree-sitter grammars (Python, JS, TS, Rust, Go, Java, Scala, C, C++, C#, PHP, Lua, Dart):

pip install 'code-graph-rag[treesitter-full]'

With semantic code search (UniXcoder embeddings):

pip install 'code-graph-rag[semantic]'

Qdrant is the default vector store for semantic search. To use Milvus Lite, install code-graph-rag[semantic,milvus], then set CGR_VECTOR_STORE_BACKEND=milvus and MILVUS_URI=./.milvus_code_embeddings.db before indexing.

To compute embeddings on an OpenAI-compatible endpoint (OpenAI, Ollama, vLLM) instead of locally, set CGR_EMBEDDING_PROVIDER=openai with OPENAI_EMBEDDING_BASE_URL and OPENAI_EMBEDDING_MODEL; torch and transformers are then not required locally.

Prerequisites

• Python 3.12+
• Docker (for Memgraph)
• cmake (for building pymgclient)
• ripgrep (rg) (for shell command text searching)

CLI Quick Start

The package installs a cgr command.

Start Memgraph, parse a repo, and query it:

cgr daemon up                              # start Memgraph + Qdrant
cgr start --repo-path ./my-project \
          --update-graph --clean           # parse & launch interactive chat

Index to protobuf for offline use:

cgr index -o ./index-output --repo-path ./my-project

Export knowledge graph to JSON:

cgr export -o graph.json

AI-guided optimisation:

cgr optimize python --repo-path ./my-project

Find dead code (functions unreachable from any entry point):

cgr dead-code                                   # scan the indexed project
cgr dead-code -e main --exclude '*.gen.*'       # add roots, skip generated code
cgr dead-code --format json --fail-on-found     # CI-friendly report

Results are candidates for review, not a guaranteed delete list. See the Dead Code Detection guide.

Run as an MCP server (for Claude Code):

cgr mcp-server

Check your setup:

cgr doctor

Python SDK

The cgr package provides short imports for programmatic use.

Load and query an exported graph

from cgr import load_graph

graph = load_graph("graph.json")
print(graph.summary())

functions = graph.find_nodes_by_label("Function")
for fn in functions[:5]:
    rels = graph.get_relationships_for_node(fn.node_id)
    print(f"{fn.properties['name']}: {len(rels)} relationships")

Query Memgraph with Cypher

from cgr import MemgraphIngestor

with MemgraphIngestor(host="localhost", port=7687) as db:
    rows = db.fetch_all("MATCH (f:Function) RETURN f.name LIMIT 10")
    for row in rows:
        print(row)

Generate Cypher from natural language

import asyncio
from cgr import CypherGenerator

async def main():
    gen = CypherGenerator()
    cypher = await gen.generate("Find all classes that inherit from BaseModel")
    print(cypher)

asyncio.run(main())

Semantic code search

Requires the semantic extra.

from cgr import embed_code

embedding = embed_code("def authenticate(user, password): ...")
print(f"Embedding dimension: {len(embedding)}")

Configuration

from cgr import settings

settings.set_orchestrator("openai", "gpt-4o", api_key="sk-...")
settings.set_cypher("google", "gemini-2.5-flash", api_key="your-key")

Environment Variables

Configure via .env or environment variables:

Documentation

Full documentation, architecture details, and contribution guide: docs.code-graph-rag.com

License

MIT