IGraph

Code Knowledge Graph Builder

Parse → Summarize → Vectorize — turn your codebase into a queryable knowledge graph so AI assistants truly understand your code

Quick Start →CLI Reference

$npm install -g igraph-cli

Comparison

Same task, two outcomes

When an AI assistant connects to a code knowledge graph, it goes from "guessing" to "precise cross-modal orchestration".

Add a payment status filter to the order list.

Coding Agent

grep "order" · grep "payment status" — repeatedly searches keywords to find files

Looks like I need to add a payment status filter to the order list page. Let me find the page and query function.

Edit order list component — add payment status dropdown

Edit getOrderList — guesses a status query param

😊Done! Added payment status filter dropdown and query param.

I guessed the field name and status values based on common patterns — please verify.

⚠ Missed these

✕ PRD Acceptance Criteria — didn't read PRD-012, missed rules like "auto-close unpaid orders after 30 min"

✕ Actual DB Field — guessed status, but it's t_order.pay_status (different enum values)

✕ Status Dropdown Source — didn't use useDict dictionary, options hardcoded or missing

✕ Export Feature — exportOrderList not updated with payment status filter

✕ Backend API Params — server query missing pay_status filter

✕ Unit Tests — test cases will fail due to param structure change

Coding Agent✦ IGraph MCP

igraph_explore "order list payment status filter" · semantic search (not keyword grep)

🎯 Requirements clarified: one query hits PRD-012 + getOrderList + pay_status field

PRD PRD-012 "Order Management" — unpaid over 30 min not displayed, etc.

FN getOrderList — order list query function

DB t_order.pay_status — actual payment status field

FN useDict — payment status dropdown dictionary

FN exportOrderList — order export function

Requirements clarified: filter by pay_status field, status values from useDict dictionary, unpaid orders over 30 min not displayed. Cross-modal change plan:

🚀 Plan · 6 steps

1. Align with PRD-012 acceptance rules (hide timeout unpaid)

2. Add pay_status filter to getOrderList

3. Add pay_status param to backend API

4. Integrate useDict payment status dictionary

5. Sync pay_status filter to exportOrderList

6. Update order list unit tests

Edit precise cross-modal changes across 6 files

😎Done. Exact field names, PRD-compliant rules, export & tests updated — full diff ready.

✅ Nothing missed

✓ PRD acceptance criteria✓ Actual DB field✓ Status dropdown source✓ Export feature✓ Backend API params✓ Unit tests

Visual Exploration

See the full picture of your codebase

Every symbol is a node, every call and dependency is an edge. Larger nodes have more connections.

igraph · graph view

Code Symbols PRD Docs DB Schema Calls & Dependencies

Workflow

Three commands, zero to queryable

Init, build, register. IGraph handles parsing, summarization, vectorization, and MCP integration automatically.

Terminal

❯ igraph init

✓ Config file generated:.igraph/config.json

❯ igraph build

✓ Parsed:142 files,856 symbol nodes

✓ Semantic summaries:856 symbols + 142 files

✓ Vector index:998 embeddings written

took 12.3s

❯ igraph register

✓ Registered to Claude Code：.mcp.json

✓ Registered to Cursor：.cursor/mcp.json

Core Capabilities

A knowledge graph built for AI coding

From code parsing to semantic understanding, from vector retrieval to MCP integration — covering the full pipeline.

🌳

Multi-Language Parsing

tree-sitter powered 5-pass pipeline, supports TypeScript / JavaScript / Python, extracts functions, classes, components and call relationships

🧠

Semantic Summaries

LLM generates semantic summaries for files and symbols; automatically falls back to heuristics without an API Key

🔍

Dual-Channel Retrieval

BGE-M3 vectors + FTS5 full-text search fused via RRF for precise code symbol and logic retrieval

📎

Multimodal Mounting

Mount PRD docs, DB schemas and other external resources; build cross-modal associations by semantic similarity

⚡

Incremental Build

SHA-256 diff-based cascading update mechanism — only rebuilds affected files, dramatically reducing build time

🤖

MCP Integration

One command to connect with Cursor / Claude Code — AI assistants can directly query the knowledge graph

How It Works

From source code to intelligent retrieval

IGraph transforms a codebase into an AI-queryable knowledge graph in three phases.

Parse — Build Structure Graph

tree-sitter parses source code, extracts symbols like functions, classes, and components, builds call, inheritance, and import relationship edges.

Summarize — Inject Understanding

LLM generates semantic summaries for each symbol and file; BGE-M3 converts summaries and source into 1024-dim vectors.

Serve — AI Assistants Query Directly

Exposes 4 read-only retrieval tools via MCP Server. AI assistants ask in natural language, get relevant symbols and context.

Let AI assistants understand your entire codebase

Three commands to integrate — no more lost context.

Get Started →