Code-Graph-RAG: Analyzing Monorepo Codebases and Call Chains with Knowledge Graphs Instead of Flat Text Search

An in-depth look at Code-Graph-RAG, an open-source tool using Tree-sitter and Memgraph to turn multi-language codebases into knowledge graphs, converting natura

tau · October 3, 2026

#Code-Graph-RAG #Memgraph #Tree-sitter #MCP #KnowledgeGraph #CodeAnalysis #Monorepo

Code-Graph-RAG: Analyzing Monorepo Codebases and Call Chains with Knowledge Graphs Instead of Flat Text Search

When collaborating with AI coding agents across large monorepos or multi-language codebases, one of the most persistent failure modes is accurately answering the question: "If I modify this function, what downstream components are impacted?" Standard LLM workflows rely heavily on flat text grep or file chunk search. This frequently introduces noise by mixing up identical identifier names in comments and test suites, or misses actual callers residing in separate packages.

Code-Graph-RAG codebase knowledge graph architecture and Tree-sitter parsing workflow in Memgraph

Image source: vitali87 / GitHub

The open-source project Code-Graph-RAG (vitali87/code-graph-rag) addresses these limitations by parsing entire repositories into an interconnected knowledge graph stored in Memgraph. By translating plain English queries into Cypher graph traversals, it transforms dependency tracking from statistical guessing into deterministic graph walks, while offering AST-based editing, diff previews, and MCP (Model Context Protocol) server integration for agents like Claude Code.

Multi-Language Knowledge Graph Ingestion with Tree-sitter and Compiler Frontends

The core architecture of Code-Graph-RAG treats codebases as structured relational graphs rather than flat strings of text.

  • Tree-sitter AST Parsing: A robust, language-agnostic Tree-sitter parser scans source files to extract functions, classes, methods, and modules along with their definitions, call sites, inheritance hierarchies, and import dependencies.
  • Compiler-Grade Frontend Integration: For deeper semantic clarity and type resolution, the parser can optionally plug into compiler-grade frontends where available, including libclang for C/C++, go/types for Go, javac for Java, Roslyn for C#, and Jedi for Python.
  • Unified Schema in Memgraph: Extracted nodes and edges are ingested into Memgraph, a high-performance in-memory graph database, under a standardized language-neutral schema. This provides unified structural visibility across multi-language monorepos within a single graph.

Natural Language Cypher Queries and Exact Blast Radius Analysis

Once the knowledge graph is populated, developers and AI agents can query the codebase using natural language.

  • Natural Language to Cypher Translation: Queries such as "What callers depend on this function?" are translated into Cypher graph queries. Rather than guessing via keyword relevance, the engine executes a graph traversal along dependency edges to return exact call chains.
  • Eliminating Noise and False Positives: Because identifiers are anchored to specific AST nodes and namespaces, homonymous functions or variables in unrelated scopes are kept strictly separate.
  • Dead Code Detection: Isolated nodes without inbound call edges can be identified immediately, simplifying the cleanup of obsolete functions and unreachable code paths.

AST-Based Diff Edits, ast-grep, Runtime Traces, and MCP Integration

Code-Graph-RAG extends beyond read-only navigation by supporting verified code modifications and runtime data integration.

  • AST-Level Modifications with Diff Previews: Instead of blindly overwriting source files, the system computes structured AST modifications and presents diff previews for verification before applying changes.
  • Structural Rewriting with ast-grep: It integrates ast-grep for syntax-aware pattern search and replacement, allowing precise refactoring across repositories.
  • Runtime Trace Merging (cgr trace): Dynamic dispatches and execution flows that static analysis cannot fully capture can be recorded via cgr trace and merged directly into the existing knowledge graph.
  • MCP (Model Context Protocol) Server Support: Coding assistants like Claude Code can connect to Code-Graph-RAG as an MCP server, inspecting dependencies and applying modifications through structured tool calls.
  • Support for 13+ Languages: The platform fully supports multi-language projects including Python, TypeScript, JavaScript, Rust, Go, Java, C, C++, C#, PHP, Lua, and Dart.

Practical Considerations and Index Hygiene

When integrating graph-based code analysis into production workflows, several operational factors should be considered:

  • Periodic Re-Indexing: If the graph database is not updated following code modifications, queries may return stale call chains. Teams should incorporate re-indexing into local hooks or CI workflows.
  • FFI and Cross-Language Boundary Configuration: Complex cross-language boundaries, such as Python calling C++ native extensions via FFI bindings, may require tailored analysis configuration to ensure edges remain unbroken across language frontiers.

Sources