🧭 Code vs. documents: when to stop at Graphify

When you point Graphify at a code repository, it doesn’t “read” the code as continuous text. It structural parsing: reads the language syntax and builds a tree of functions, classes, modules, and their calls. The result becomes the graph: nodes are functions and modules, and edges are "calls," "imports," and "defines."

The practical consequence is important: extracting a code base is deterministic, fast, and no API key. There’s no possibility of “hallucination”—the tree reflects the file’s syntax exactly. You can run it on a huge project without spending a model token.

# extrai a AST do código em ./src — tree-sitter, sem API key
graphify extract ./src

# pergunta direto ao grafo de código já extraído
graphify query "quais funções chamam autenticar()?"

↑ graphify extract is the path headless (plain terminal). For code, it runs without any key—the tree-sitter does all the work.

Documents don't have program syntax — a docs page or a PDF is prose. So Graphify changes its strategy: instead of tree-sitter, it uses a LLM to read each document and understand what concepts it covers and how they relate to one another. It’s extraction semantics: by meaning, not by form.

↑ Same Graphify, two engines. Code becomes structure (AST, no key); documents become meaning (LLM, semantic). Inside the /graphify in Claude Code, the session already provides the model—you don’t need a key here either.

There’s a temptation to move every graph into Obsidian “because you can.” But the graph is already useful alone. O graph.html (self-contained interactive visualization) opens in the browser without a server, and graphify query answers questions directly. To explore a specific codebase or get a quick answer, this enough.

↑ Arrows green = yes, red = no. The question is never “can it be exported?”—it’s “does this knowledge need to live outside the terminal?”

Obsidian comes in when knowledge stops being a one-off lookup and becomes collection. When you export, Graphify generates one markdown file per node with wikilinks [[nome-do-no]] and backlinks — then this graph can coexist with your other notes, projects, and ideas in a single navigable vault.

# exporta o grafo como vault Obsidian (um .md por nó + backlinks)
/graphify ./claude-code-docs --obsidian --obsidian-dir ~/vault/graphify/claude-code

↑ The flag --obsidian only exists along the path skill (/graphify inside Claude Code), not in the graphify extract headless.

Code and docs change. A graph frozen on the day of extraction grows stale and misleads. The solution isn't to re-extract everything from scratch every time — it's incremental update: Graphify stores a cache/ per file (by content hash) and only reprocesses what changed.

# reprocessa só os arquivos modificados desde a última vez
graphify update .

# rebuild automático: fica observando e atualiza ao salvar
graphify watch .

↑ graphify update . = on-demand update; graphify watch . = continuous. Inside Claude Code, the equivalent is /graphify . --update e /graphify . --watch.

All this structure comes at a price — and it’s real, not just hype. Choosing wisely between code and documents, stopping at the graph or going to the vault, is a account of three variables: extraction time/cost, graph noise, and maintenance effort. Weighing them first keeps you from becoming a collector of abandoned graphs.