graphify-rs

AI-powered knowledge graph builder — transform code, docs, papers, and images into queryable, interactive knowledge graphs.

中文文档 | CLI Reference | Changelog

What is graphify-rs?

graphify-rs is built around Andrej Karpathy's /raw folder workflow: drop anything into a folder — papers, tweets, screenshots, code, notes — and get a structured knowledge graph that shows you what you didn't know was connected.

It is a Rust rewrite of graphify (Python), with full feature parity and significant performance improvements.

Three things it does that an LLM alone cannot

1. Persistent graph — relationships are stored in graph.json and survive across sessions. Ask questions weeks later without re-reading everything.
2. Honest audit trail — every edge is tagged EXTRACTED, INFERRED, or AMBIGUOUS. You always know what was found in source vs. what was inferred.
3. Cross-document surprise — community detection finds connections between concepts in different files that you would never think to ask about directly.

Use cases

• New to a codebase — understand architecture before touching anything
• Research corpus — papers + tweets + notes → one navigable graph with citation + concept edges
• Personal /raw folder — drop everything in, let it grow, query it anytime
• Agentic workflows — AI agents query the graph via MCP server for grounded, structured context

Compared to the Python Version

Output format is fully compatible — graph.json uses the same NetworkX node_link_data schema.

Installation

From crates.io

cargo install graphify-rs

From source

git clone https://github.com/TtTRz/graphify-rs.git
cd graphify-rs
cargo install --path .

Quick Start

graphify-rs build                    # build knowledge graph from current directory
open graphify-out/graph.html         # explore in browser
graphify-rs query "how does auth work?"  # query the graph

For the full CLI reference, see docs/CLI.md.

How It Works

Pipeline Overview

 Source Files          graphify-rs build
 ┌──────────┐    ┌─────────────────────────────────────────────────────────┐
 │ .py .rs  │    │                                                         │
 │ .go .ts  │───▶│  detect → extract → build → cluster → analyze → export │
 │ .md .pdf │    │                                                         │
 │ .png     │    └──────────┬──────────────────────────────────────────────┘
 └──────────┘               │
                            ▼
                  graphify-out/
                  ├── graph.json        (queryable graph data)
                  ├── graph.html        (interactive visualization)
                  ├── GRAPH_REPORT.md   (analysis report)
                  ├── wiki/             (per-community wiki pages)
                  └── obsidian/         (Obsidian vault)

Two-Pass Extraction

Pass 1 — Deterministic AST extraction (free, fast, always runs):

Uses tree-sitter to parse source code into ASTs, then extracts functions, classes, imports, and call relationships. Supports 21 languages with 11 native tree-sitter grammars and regex fallback for the rest. Every edge from this pass is tagged EXTRACTED with confidence 1.0.

Pass 2 — Semantic extraction via Claude API (optional, --no-llm to skip):

Sends document/paper/image content to the Claude API to discover higher-level relationships that syntax alone cannot reveal — conceptual links, shared assumptions, design rationale. Edges from this pass are tagged INFERRED with confidence scores from 0.4 to 0.9.

Confidence System

Every edge in the graph carries a confidence tag:

This ensures you always know which relationships are facts vs. guesses.

Leiden Community Detection

After building the graph, graphify-rs runs the Leiden algorithm to partition nodes into communities:

1. Louvain phase — greedy modularity optimization, moving nodes to neighboring communities for maximum modularity gain
2. Refinement phase — BFS within each community to ensure internal connectivity; disconnected sub-communities are split
3. Small community merging — communities with < 5 nodes are merged into their most-connected neighbor

Each community receives a cohesion score (ratio of actual intra-community edges to maximum possible), and the report surfaces "god nodes" (highest-degree hubs) and "surprising connections" (edges that bridge different communities).

Architecture

14 crates organized as a Cargo workspace:

Output Formats

CLI Reference

See docs/CLI.md for the complete command reference with all flags, defaults, and examples.

Quick overview:

graphify-rs build [--path .] [--no-llm] [--format json,html]  # build graph
graphify-rs query "question" [--dfs] [--budget 2000]           # query graph
graphify-rs watch --path .                                      # auto-rebuild
graphify-rs serve                                                # MCP server
graphify-rs diff old.json new.json                              # compare graphs
graphify-rs stats graph.json                                    # show statistics

Agent Integration

graphify-rs integrates with AI coding agents (Claude Code, Codex, OpenCode, etc.) via skill installation and MCP server.

graphify-rs install                # install skill globally
graphify-rs claude install         # project-level: CLAUDE.md + PreToolUse hook
graphify-rs serve                  # start MCP server for agent queries

Once installed, agents automatically check the knowledge graph before answering architecture questions and rebuild it after code changes.

For full agent setup instructions, see the Agent Integration section of the CLI reference.

MCP Server Tools

Supported Languages (21)

Contributing

See CONTRIBUTING.md for development setup, code style, and PR guidelines.

License

MIT — see LICENSE.

This project is a Rust rewrite of graphify by safishamsi.