Knowledge Graph Builder

# CLAUDE.md This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository. ## Project Overview KGB-mcp is a Knowledge Graph Builder MCP (Model Context Protocol) server that transforms text or web content into structured knowledge graphs using AI-powered entity extraction and relationship mapping. The project is built as a Gradio application designed for the MCP Hackathon 2025. ## Core Architecture - **Entry Point**: `app.py` - Main application file containing the Gradio interface and MCP server - **AI Model**: Uses Mistral AI (`mistralai/Mistral-Small-24B-Instruct-2501`) via HuggingFace Inference Client - **Web Scraping**: BeautifulSoup for extracting text content from URLs - **Output Format**: Structured JSON knowledge graphs containing entities and relationships ## Key Functions - `extract_text_from_url()` - Scrapes and cleans text from web URLs (app.py:15) - `extract_entities_and_relationships()` - Uses Mistral AI to extract structured knowledge graphs (app.py:42) - `build_knowledge_graph()` - Main orchestration function that handles both text and URL inputs (app.py:134) ## Environment Setup **Required Environment Variables:** - `HF_TOKEN`: HuggingFace API token for accessing Mistral AI through the inference client **Dependencies Installation:** ```bash pip install -r requirements.txt ``` ## Running the Application **Development:** ```bash python app.py ``` The application launches a Gradio interface with MCP server capabilities enabled (`mcp_server=True`). ## Input/Output Format **Input**: Text content or web URLs **Output**: JSON structure containing: - `source`: Information about the input (type, value, content preview) - `knowledge_graph`: Extracted entities and relationships with counts - `metadata`: Model information and content length **Entity Types**: PERSON, ORGANIZATION, LOCATION, CONCEPT, EVENT, OTHER **Relationship Types**: Custom relationship types extracted by the AI model ## Content Limits - URL content: Limited to first 5000 characters - AI analysis: Uses first 3000 characters of content - Content preview: First 200 characters in output ## Error Handling The application includes comprehensive error handling for: - Invalid URLs or network failures - Missing API tokens - JSON parsing errors from LLM responses - Malformed or empty inputs