Semantic Summaries

IGraph generates LLM-powered semantic summaries for files and symbols, so the graph contains not just structural information but also semantic understanding.

How It Works

During the build process, IGraph sends each symbol's source code and context to an LLM to generate concise semantic descriptions. Summaries are stored in the graph database for retrieval and display.

Symbol Summaries

For each function, class, component, etc., the LLM generates a one-line summary describing its responsibility and behavior:

verifyToken — Validates JWT token authenticity, checks signature and expiration, returns decoded user info

File Summaries

For each file, a file-level overview is generated based on its exported symbols and internal logic:

src/auth/jwt.ts — JWT authentication module providing token generation, validation, and refresh capabilities

Heuristic Fallback

When there's no IGRAPH_API_KEY or the --no-llm flag is used, IGraph automatically switches to heuristic summaries:

igraph build --no-llm

The heuristic approach generates descriptions from symbol names, parameter signatures, and code structure without network access:

verifyToken(token: string): Promise<User> — Accepts a token parameter, returns a Promise of User type

Less precise than LLM summaries, but guarantees offline availability.

Configuration

In the llm section of .igraph/config.json:

{
  "llm": {
    "baseURL": "https://api.openai.com/v1",
    "model": "gpt-4o-mini",
    "fileSummaryModel": "gpt-4o",
    "temperature": 0,
    "maxConcurrency": 5,
    "promptVersion": "v1.0"
  }
}

Field	Description
model	Model for symbol summaries (lightweight, fast)
fileSummaryModel	Model for file summaries (stronger comprehension)
maxConcurrency	Concurrent request limit to avoid rate limiting

Cost Control

Lightweight models like gpt-4o-mini are sufficient for symbol summaries. Use a stronger model for file summaries to get better overviews.