ODRL Multi-Agent LLM

Transform natural language into validated ODRL policies using multi-agent AI

Automated generation of machine-readable data usage policies through a four-agent pipeline with human-in-the-loop validation.

Key Features

• 4-Stage Pipeline: Parser → Reasoner → Generator → Validator
• Dual Checkpoints: Semantic conflict detection + SHACL validation
• Iterative Refinement: Auto-corrects validation errors (not single-shot)
• Multi-Model Support: Groq, Ollama, OpenAI-compatible endpoints
• Interactive UI: Real-time monitoring, manual/auto execution modes

Quick Start

Local Development:

# Backend
cd backend && uv sync
uv run uvicorn main:app --reload --host 0.0.0.0 --port 8000

# Frontend
cd frontend && npm install && npm start

Production: 📖 Deployment Guide →

Architecture

graph LR
    A[Natural Language] --> B[Parser Agent]
    B --> C{Checkpoint I}
    C -->|✓ Valid| D[Reasoner Agent]
    C -->|✗ Conflicts| E[User Review]
    D --> F[Generator Agent]
    F --> G{Checkpoint II}
    G -->|✓ Valid| H[ODRL Policy]
    G -->|✗ Invalid| F

Loading

Workflow Diagram

Figure: Multi-agent pipeline with dual human checkpoints. Reasoner (Checkpoint I) enables pre-generation review; Validator (Checkpoint II) enables post-generation refinement. Red dashed: edit input; orange dashed: regenerate; green: continue. Supports per-agent LLM configuration.

Documentation

Guide	Description
Deployment	Ubuntu production setup
Configuration	Environment & settings
Development	Local setup & structure
Testing	Test procedures
API Reference	Backend endpoints

🎬 Demo

Video Demo: [Watch on YouTube](coming soon...)

Screenshot Demos: Check the full demos here.

Tech Stack

Backend: FastAPI • LangChain • LangGraph • RDFLib • PySHACL
Frontend: React • Tailwind CSS
LLMs: Groq • Ollama • OpenAI-compatible • Google GenAI

Research Context

Developed at Fraunhofer FIT & RWTH Aachen University for:

• NFDI4Culture
• Daten-Raum-Kultur (DRK)
• Cultural Heritage Dataspaces

Authors

• Daham M. Mustafa - Fraunhofer FIT, Sankt Augustin, Germany
• Diego Collarana - Fraunhofer FIT, Sankt Augustin, Germany
• Christoph Lange - Fraunhofer FIT & RWTH Aachen University, Germany
• Christoph Quix - Fraunhofer FIT & RWTH Aachen University, Germany
• Stefan Decker - Fraunhofer FIT & RWTH Aachen University, Germany

Citation

If you use this software in your research, please cite:

BibTeX:

@software{mustafa_2025_odrl,
  author       = {Mustafa, Daham M. and
                  Collarana, Diego and
                  Lange, Christoph and
                  Quix, Christoph and
                  Decker, Stefan},
  title        = {ODRL Multi-Agent LLM: A Multi-Agent System for 
                  ODRL Policy Generation},
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v2.1.0},
  doi          = {10.5281/zenodo.17670391},
  url          = {https://doi.org/10.5281/zenodo.17670391}
}

Contributing

Contributions welcome! See CONTRIBUTING.md

License

MIT License - see LICENSE

🔗 Links

• Repository: github.com/Daham-Mustaf/odrl-multi-agent-llm
• Issues: Report a bug
• ODRL Spec: w3.org/TR/odrl-model