A sequential multi-agent fact-checking system that grounds claims in evidence. Built with LangGraph, OpenAI, and Tavily.
GroundCrew automatically fact-checks text by:
- Extracting factual claims from input
- Searching for evidence via web search
- Verifying claims against evidence
- Reporting findings with confidence scores
# Install
poetry install
# Configure API keys
cp env.template .env
# Add your OpenAI and Tavily keys to .env
# Run with text
poetry shell
python main.py "Your text to fact-check"
# Or run with URL (requires FIRECRAWL_API_KEY)
python main.py --url https://example.com/articlefrom groundcrew.workflow import run_fact_check
result = run_fact_check(
input_text="The Eiffel Tower is 330 meters tall.",
openai_api_key="sk-...",
tavily_api_key="tvly-..."
)
print(result.final_report)# Fact-check a web page
python main.py --url https://example.com/article
# With custom output and model
python main.py -u https://example.com/article -o report.md --model gpt-4
# Combine with Wikipedia-only mode
python main.py --url https://example.com/article --wikipedia-onlyComplete documentation is available in the wiki:
- Getting Started - Installation and first steps
- Usage Guide - CLI and Python API usage
- Architecture - System design and how it works
- API Reference - Models and configuration
- Examples - Code examples and use cases
- Development - Contributing and extending
- Troubleshooting - Common issues and solutions
- ✅ Sequential 4-stage pipeline using LangGraph
- ✅ Structured output with Pydantic validation
- ✅ Type-safe throughout with full type hints
- ✅ Real-time web search via Tavily
- ✅ URL scraping with Firecrawl integration
- ✅ Configurable quality presets
- ✅ Comprehensive test coverage
Evaluated on FEVER dataset (100 samples, GPT-4o):
| Configuration | Overall | SUPPORTS | REFUTES | NOT ENOUGH INFO |
|---|---|---|---|---|
| Web Search | 71% | 88% | 82% | 42% |
| Wikipedia-only | 72% | 91% | 88% | 36% |
See evals/ for evaluation scripts and detailed results.
- LangGraph - Workflow orchestration
- LangChain - LLM framework
- OpenAI GPT - Language models
- Tavily - Search API
- Pydantic - Data validation
- Poetry - Dependency management
- Python 3.12.6
- OpenAI API key (get one)
- Tavily API key (get one)
- Firecrawl API key (get one) - optional, required for URL scraping
MIT License - see LICENSE file
Powered by modern AI tools: LangGraph, LangChain, OpenAI, and Tavily.