code-graph-rag

Code‑Graph‑RAG is a command‑line utility for building and interrogating a knowledge graph from a multi‑language codebase. It uses Tree‑sitter to parse source files into an abstract syntax tree, then stores the resulting graph in Memgraph. The AI agent translates plain‑English prompts into Cypher queries that run against that graph, enabling you to ask questions such as “Which functions handle authentication?” and instantly receive a list of file paths, function names and line numbers. The output is a concise, machine‑readable list that can be copied into a report or entered into a version control diff.

The tool is especially useful for developers who need an overview of a codebase that spans several languages. It is not intended for simple single‑file scripts or for repositories that use custom, non‑Tree‑sitter grammars unless you supply the pattern files yourself. If your project lives entirely in a language with a quality Tree‑sitter grammar, such as Python, JavaScript, Rust, Java, C#, or Go, then you can rely on the unified graph schema that treats all entities the same way and supports cross‑lingual queries.

Editing is performed on the graph nodes: you can target a specific function, verify its definition line number and provide a code snippet to replace the body. The replacement is validated against the AST, then a visual diff preview is shown before the change is committed. This means the tool does not merely perform textual substitutions; it respects scopes, imports and language syntax. For larger refactors the tooling integrates with `ast‑grep` to locate pattern‑based matches and perform structure‑aware find‑and‑replace across the entire repository. It does not, however, support regular expression searches on raw file content.

Data‑flow tracing builds FLOWS_TO edges that connect assignments, function calls and returns to I/O sinks, giving you end‑to‑end provenance. Dead‑code detection walks reference edges from defined entry points and marks unreachable functions; the output can be integrated into CI pipelines with a failure mode that asserts no uncovered functions exist. These static analyses are available for all 14 built‑in languages and can be extended with custom grammars if you add a single ast‑grep pattern file.

In a typical workflow you first run `cgr daemon up` to start Memgraph and Qdrant, optionally provisioning them for an on‑prem or air‑gapped environment.