graphify（graphify 文档）

Graphify Skill

Graphify turns a folder of code, docs, images, and videos into a queryable knowledge graph. After running graphify ., the generated GRAPH_REPORT.md is referenced by the assistant when you ask it to explore the codebase. Use it to navigate unfamiliar codebases, trace architectural intent, and share context efficiently.

Credential note: Semantic extraction (docs, PDFs, images) calls an external LLM using your own API key (ANTHROPIC_API_KEY or OPENAI_API_KEY). AST code extraction and Whisper transcription run fully locally with no API key required.

Token efficiency: Reading GRAPH_REPORT.md is ~71.5× cheaper than reading raw source files. Always check it first.

Installation

Minimum install

pip install graphify

Recommended install with all extras

pip install "graphify[pdf,video,watch,svg]"

Register with OpenClaw

graphify claw install

This copies the skill into OpenClaw's global skill directory and writes an AGENTS.md to the project root so the graph is consulted on every tool call.

Register with other platforms

graphify claude install     # Claude Code (CLAUDE.md + PreToolUse hook)
graphify cursor install     # Cursor (.cursorrules)
graphify codex install      # Codex (AGENTS.md)
graphify copilot install    # GitHub Copilot CLI
graphify gemini install     # Gemini CLI
graphify aider install      # Aider

Git post-commit hook (optional, local project only)

Installs a hook in the current project's .git/hooks/ — no global changes.

graphify hook install

Building a Knowledge Graph

First run

graphify .

Graphify runs a three-pass pipeline:

1. AST extraction — deterministic tree-sitter parsing of all code files (no LLM calls)
2. Transcription — local Whisper processing of any video/audio
3. Semantic extraction — parallel LLM calls (using your configured API key) analyze docs, papers, and images

Incremental update (changed files only)

graphify . --update

Uses a SHA-256 cache in graphify-out/cache/ — safe to run after every save.

Watch mode (auto-sync as files change)

graphify . --watch

Deep mode (aggressive inferred-edge extraction)

graphify . --mode deep

Adds INFERRED edges with confidence scores (0.0–1.0) and AMBIGUOUS edges flagged for review.

Preserve edge directionality

graphify . --directed

Re-cluster without re-extracting

graphify . --cluster-only

Skip HTML visualization (faster CI runs)

graphify . --no-viz

Output Artifacts

All artifacts land in graphify-out/:

Additional export formats

graphify . --svg        # SVG visualization
graphify . --graphml    # Gephi / yEd compatible export
graphify . --wiki       # Wikipedia-style article per node

Querying the Graph

# Natural-language semantic search
graphify query "where is authentication handled?"

# Trace a specific path (DFS traversal)
graphify query "how does the request reach the database?" --dfs

# Shortest path between two nodes
graphify path "AuthMiddleware" "PostgresAdapter"

# Plain-language explanation of a node
graphify explain "UserSessionManager"

Adding External Content

graphify add <arxiv-url>     # Fetch and index a research paper
graphify add <x.com-url>     # Fetch and index a tweet
graphify add <video-url>     # Download and transcribe video/audio

After adding, run graphify . --update to integrate the new nodes into the graph.

Ignoring Files

Create .graphifyignore in the project root (same syntax as .gitignore):

node_modules/
dist/
build/
.next/
vendor/
*.generated.*
*.min.js
*.lock
graphify-out/cache/

See templates/graphifyignore.txt in this skill for a comprehensive starter.

Relationship Types

Use --mode deep to maximize INFERRED coverage. Filter AMBIGUOUS edges when precision matters.

Best Practices for OpenClaw / Claude Code

1. Always read the report before exploring

Read graphify-out/GRAPH_REPORT.md

The report identifies "god nodes" (highest-degree concepts), community clusters, and suggested entry-point questions. Use these to focus subsequent searches.

2. Map communities to features

Each community in the report corresponds to a functional area of the codebase. When fixing a bug or adding a feature, identify the relevant community first, then limit file reads to that cluster.

3. Use graphify query before grep

For open-ended questions ("where is X handled?"), prefer graphify query — it searches the semantic graph and returns ranked node paths rather than raw text matches.

4. Use graphify path for impact analysis

Before editing a node, run graphify path "NodeA" "NodeB" to find coupling chains and assess blast radius.

5. Use --update aggressively

After any significant edit, run graphify . --update to keep the graph current. The cache makes this cheap.

6. Team workflows

• Commit graphify-out/ (excluding graphify-out/cache/) so teammates get graph context immediately on checkout.
• One team member builds the initial graph; all others benefit without LLM cost.
• Install graphify hook install to auto-rebuild on every commit.

7. Multimodal context

Place architecture diagrams, whiteboard photos, or design mockups in the project directory before running graphify .. Claude vision will link visual concepts to code nodes.

Supported Languages

Python, JavaScript, TypeScript, Go, Rust, Java, C, C++, Ruby, C#, Kotlin, Scala, PHP, Swift, Lua, Zig, PowerShell, Elixir, Objective-C, Julia, Verilog, SystemVerilog, Vue, Svelte, Dart

Plus: Markdown, MDX, HTML, plain text, RST, PDF, PNG/JPG/WebP/GIF images, MP4/MOV/MKV/WebM/MP3/WAV/M4A/OGG media.

Troubleshooting