code-graph-rag

Product Introduction

1. Definition: Code-Graph-RAG is an open-source, AI-powered code intelligence platform that combines Retrieval-Augmented Generation (RAG) with a semantic knowledge graph. It is a developer tool for static code analysis, querying, and automated refactoring.
2. Core Value Proposition: It exists to solve the problem of understanding and navigating complex, multi-language monorepos by automatically building a comprehensive, queryable knowledge graph from source code, enabling developers to ask questions and perform edits in natural language.

Main Features

1. Multi-Language Knowledge Graph Construction: The platform parses source code across 14 programming languages (Python, TypeScript, Java, C++, Go, etc.) using Tree-sitter to generate Abstract Syntax Trees (ASTs). It then transforms these ASTs into a unified graph schema stored in Memgraph, creating nodes for functions, classes, modules, and edges for relationships like calls, imports, and data flow (FLOWS_TO).
2. Natural Language to Cypher Query Translation: Users can ask questions like "What functions handle authentication?" in plain English. An integrated AI model (e.g., Google's Gemini) translates this intent into a precise Cypher query, which is executed against the Memgraph graph database to return structured results about code location and relationships.
3. AST-Aware Surgical Code Editing & Structural Search: Beyond querying, it enables AI-driven code modifications. The surgical_replace_code tool targets exact AST blocks (e.g., a specific function body) for replacement, providing a visual diff. The structural_search and structural_replace tools use ast-grep for pattern matching on code structure, enabling safe, syntax-aware refactoring across an entire codebase.
4. Model Context Protocol (MCP) Server Integration: Code-Graph-RAG functions as an MCP server, providing 15 tools (indexing, querying, editing, searching) directly within AI coding assistants like Claude Code. This allows developers to interact with their code graph without leaving their IDE or chat interface.
5. Advanced Code Analysis (Data-Flow & Dead Code): It performs sophisticated static analysis. Data-flow tracing uses FLOWS_TO edges to track values from sources to sinks for security and provenance. Dead code detection performs graph reachability analysis from entry points to identify functions and methods that are never called.

Problems Solved

1. Pain Point: The overwhelming complexity and scale of modern, polyglot monorepositories make manual code navigation, impact analysis, and systematic refactoring time-consuming and error-prone.
2. Target Audience: Senior Software Engineers, Engineering Managers, and Platform/DevEx teams working in large, multi-language codebases; Security Engineers performing code audits; and Developers using AI coding assistants (Claude Code) who need deeper codebase context.
3. Use Cases: Onboarding new engineers to a massive codebase; performing large-scale refactoring or library migrations safely; identifying security vulnerabilities through data-flow analysis; cleaning up dead code before a release; and answering architectural questions about dependencies and call chains.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Unifies code analysis across 14 languages into a single graph model. Deep IDE integration via MCP provides a seamless workflow. Goes beyond text search with AST-level understanding for precise edits. Open-source core allows for customization and inspection.
   • Cons: Requires running local services (Memgraph, Qdrant), adding infrastructure overhead. The accuracy of natural language-to-Cypher translation depends on the configured AI model (Gemini, OpenAI, etc.). Primarily focused on static analysis, lacking runtime data.

2. Key Alternatives & Differentiation:

   • Sourcegraph Cody: A competitor offering AI-powered code search and chat. Code-Graph-RAG differentiates by being open-source, offering an on-premise deployment model, and its foundational knowledge graph enabling complex relational queries and AST-level edits that pure chat interfaces cannot perform.
   • Traditional Static Analysis Tools (SonarQube, Semgrep): These tools excel at finding bugs and enforcing rules. Code-Graph-RAG complements them by adding a holistic, queryable graph view of the entire codebase and a natural language interface for exploration and change, rather than just linting.
   • Custom Scripts (ripgrep, ast-grep): While powerful, these require deep expertise to write and maintain. Code-Graph-RAG productizes this capability with a managed graph database, an AI translation layer, and a unified toolset, lowering the barrier to advanced codebase analysis.

Frequently Asked Questions (FAQ)

1. How does Code-Graph-RAG differ from GitHub Copilot? GitHub Copilot is an AI pair programmer focused on code completion within a single file. Code-Graph-RAG is a codebase intelligence platform that builds a global knowledge graph of your entire repository, enabling architectural queries, cross-file refactoring, and deep semantic search that Copilot does not provide.
2. What is required to run Code-Graph-RAG locally? You need Python with uv, and the tool will start a local Memgraph (graph database) and Qdrant (vector database) instance via Docker. You also need an API key for an AI provider (Google AI, OpenAI, etc.) to power the natural language-to-Cypher translation.
3. Can Code-Graph-RAG handle private or enterprise codebases? Yes, its core is open-source and designed for on-premise or air-gapped deployment, ensuring code never leaves your infrastructure. The enterprise offering provides managed cloud hosting, custom development, and support contracts for regulated environments.
4. How accurate is the dead code detection feature? It uses static graph reachability analysis from declared entry points, which is highly accurate for identifying functions with no callers in the codebase. However, it may have limitations with dynamic language features or reflection, which typically require complementary dynamic analysis tools.
5. What support is available for languages not in the official 14? The system can be extended using ast-grep pattern files to add structural support for new languages. For full AST parsing, a Tree-sitter grammar would need to be integrated, which the enterprise "Custom Development" offering can provide.