CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Application Overview

This is a Text2Graph application that extracts knowledge graphs from natural language text. It's a Gradio web app that uses either OpenAI GPT-4.1-mini via Azure or Phi-3-mini-128k-instruct-graph via Hugging Face to extract entities and relationships from text, then visualizes them as interactive graphs.

Architecture

• app.py: Main Gradio application with UI components, visualization logic, and caching
• llm_graph.py: Core LLMGraph class that handles model selection and knowledge graph extraction
• cache/: Directory for caching visualization data (first example is pre-cached for performance)

Key Components

LLMGraph Class (llm_graph.py)

• Supports two model backends: Azure OpenAI (GPT-4.1-mini) and Hugging Face (Phi-3-mini-128k-instruct-graph)
• Uses LightRAG for Azure OpenAI integration
• Direct inference API calls for Hugging Face models
• Extracts structured JSON with nodes (entities) and edges (relationships)

Visualization Pipeline (app.py)

• Entity recognition visualization using spaCy's displacy
• Interactive knowledge graph using pyvis and NetworkX
• Caching system for performance optimization
• Color-coded entity types with random light colors

Environment Setup

Required environment variables:

HF_TOKEN=<huggingface_token>
HF_API_ENDPOINT=<huggingface_inference_endpoint>
AZURE_OPENAI_API_KEY=<azure_openai_key>
AZURE_OPENAI_ENDPOINT=<azure_endpoint>
AZURE_OPENAI_API_VERSION=<api_version>
AZURE_OPENAI_DEPLOYMENT=<deployment_name>
AZURE_EMBEDDING_DEPLOYMENT=<embedding_deployment>
AZURE_EMBEDDING_API_VERSION=<embedding_api_version>

Development Commands

# Install dependencies
pip install -r requirements.txt

# Run the Gradio app
python app.py

# Test model extraction directly
python llm_graph.py

Key Dependencies

• gradio: Web interface framework
• lightrag-hku: RAG framework for Azure OpenAI integration
• transformers: Hugging Face model integration
• pyvis: Interactive network visualization
• networkx: Graph data structure and algorithms
• spacy: Natural language processing and entity visualization
• openai: Azure OpenAI client

Data Flow

1. User inputs text and selects model
2. LLMGraph.extract() processes text using selected model backend
3. JSON response contains nodes (entities) and edges (relationships)
4. Visualization functions create entity highlighting and interactive graph
5. Results cached for performance (first example only)

Model Behavior

The application expects JSON output with this schema:

{
  "nodes": [{"id": "entity", "type": "broad_type", "detailed_type": "specific_type"}],
  "edges": [{"from": "entity1", "to": "entity2", "label": "relationship"}]
}

Testing and Development Notes

• No formal test suite exists; manual testing through the Gradio interface
• First example is automatically cached for performance on startup
• Cache files stored in cache/ directory as pickle files
• Working directory sample/ is cleared and recreated on each run
• GraphML files generated by LightRAG for Azure OpenAI model backend