🌟 Faithful-RAG

FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation (ACL 2025)
A novel RAG framework explicitly modeling fact-level conflicts for more faithful retrieval-augmented generation.

---

🚀 Highlights

• ✅ Fact-Level Faithfulness: Explicit modeling of conflicts between retrieved context and LLM generations.
• 🧑‍🔬 Strong Results: Outperforms previous RAG methods on context-faithfulness benchmarks.
• 💡 Easy Integration: Support for OpenAI / HuggingFace / LlamaFactory backends.
• 📊 Evaluation Tools: Built-in faithful evaluation pipeline with CoT support.

---

🛠️ Usage

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Quick Start Example

import asyncio
from datasets import load_dataset
from faithfulrag import FaithfulRAG

# Load dataset
ds = load_dataset("json", data_files="https://huggingface.co/datasets/eric-xiang/FaithfulRAG-Dataset/resolve/main/faithful-data.json", split="train")

# Initialize FaithfulRAG pipeline
rag = FaithfulRAG(
    backend_type="openai",  # or "hf", "llamafactory"
    model_name="gpt-3.5-turbo",
    similarity_model="all-MiniLM-L6-v2"
)

async def run_pipeline():
    # Generate self-consistent facts
    self_facts = await rag.get_self_facts(dataset)

    # Retrieve top contextual chunks
    topk_chunks = rag.get_topk_chunks(dataset, self_facts)

    # Generate predictions
    predictions = await rag.get_predictions(
        dataset, 
        topk_chunks,
        generation_type="normal_cot"  # or "scheduled_cot", "wo_cot"
    )

    # Evaluate results
    results = rag.evaluate(dataset, predictions, cot_format=True)
    print(f"Exact Match: {results['exact_match']:.2f}%")

asyncio.run(run_pipeline())

📖 Citation

If you find this work helpful, please consider citing us:

@misc{zhang2025faithfulragfactlevelconflictmodeling,
      title={FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation}, 
      author={Qinggang Zhang and Zhishang Xiang and Yilin Xiao and Le Wang and Junhui Li and Xinrun Wang and Jinsong Su},
      year={2025},
      eprint={2506.08938},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.08938}, 
}

📬 Contact

✉️ Email: [email protected]