GraphifyHome › Tree-sitter AST Extraction
Tree-sitter AST Extraction Across 19 Languages
The first pass of the Graphify pipeline is a deterministic Tree-sitter walk over every code file it finds. No LLM, no embeddings, no network — just AST nodes converted directly into graph nodes and edges.
Why Tree-sitter
Tree-sitter is an incremental parser generator with battle-tested grammars for every mainstream language. Graphify uses it because:
- It never calls the network. Source code never leaves the machine during the AST pass. The only outbound calls happen later, for docs/papers/images, and they send semantic descriptions — not raw source. See privacy & security.
- It's fast. Parsing is O(file size). On a modest repo the entire AST pass finishes in seconds.
- It's uniform. Every language exposes the same node/edge shape to the downstream graph builder, which means the Leiden clustering step doesn't need per-language special cases. See Leiden community detection.
Languages supported
Graphify ships grammars for 19 languages out of the box:
| Family | Languages |
|---|---|
| Scripting | Python, JavaScript, TypeScript, Ruby, PHP, Lua, PowerShell |
| Systems | Go, Rust, C, C++, Zig, Swift, Objective-C |
| JVM / .NET | Java, Kotlin, Scala, C# |
| BEAM | Elixir |
Adding a language is a matter of dropping its Tree-sitter grammar into the extractor and writing a small node-to-concept mapper — see ARCHITECTURE.md in the repo for the exact steps.
What Graphify pulls out of the AST
- Structural nodes — classes, functions, methods, modules, traits/interfaces, top-level variables.
- Call-graph edges — every resolved call site becomes a
callsedge taggedEXTRACTEDwith confidence1.0. - Import edges — module-level
importsso communities don't fragment across files that clearly belong together. - Rationale nodes — docstrings and rationale comments (
# NOTE:,# IMPORTANT:,# HACK:,# WHY:) are lifted out as separate nodes attached viarationale_foredges. This is how Graphify captures why the code was written the way it was, not just what it does.
Deterministic vs semantic
Everything the AST pass emits is marked EXTRACTED (confidence 1.0). It's not a guess — the token is in the file. The second pass, run by Claude subagents against docs/papers/images, emits INFERRED edges with confidence scores, and anything the model is unsure about is tagged AMBIGUOUS. Those tags survive all the way into graph.json so a reviewer can always separate ground truth from model judgment.