GAIR-NLP/factool

Employs a multi-tool reasoning architecture combining LLM-based claim decomposition, search APIs (Serper for web queries, ScraperAPI for academic literature), and code execution sandboxes to verify factuality across four distinct domains. Supports both claim-level and response-level accuracy metrics, with built-in integration for GPT-3.5/GPT-4 and a ChatGPT plugin interface. Also includes Halu-J, an open-source critique-based hallucination judge model, alongside multilingual evaluation benchmarks (English, Chinese, Japanese).

916 stars and 53 monthly downloads. No commits in the last 6 months. Available on PyPI.