🔥 TrustRAG: Enhancing Robustness and Trustworthiness in RAG

[Project page] [Paper]

Huichi Zhou*¹, Kin-Hei Lee*¹, Zhonghao Zhan*¹, Yue Chen², Zhenhao Li¹, Zhaoyang Wang³,Hamed Haddadi¹ Emine Yilmaz⁴

¹Imperial College London, ²Peking University, ³University of North Carolina at Chapel Hill, ⁴University College London

^*Equal Contribution

🔥 NEWS

• 2025.1.10 OpenAI API Inference Now Supported! Additionally, we have introduced a new module: Self-Assessment of Retrieval Correctness, enabling enhanced evaluation of retrieval accuracy.

🛝 Try it out!

🛠️ Installation

git clone https://github.com/HuichiZhou/TrustRAG.git

conda create -n trustrag python=3.10

conda activate trustrag

pip install lmdeploy

pip install beir

pip install nltk

pip install rouge_score

pip install timm==0.9.2

cd TrustRAG

python run.py

🙏 Acknowledgement

• Our code used the implementation of corpus-poisoning.
• The model part of our code is from Open-Prompt-Injection.
• Our code used beir benchmark.
• Our code used contriever for retrieval augmented generation (RAG).
• Our code used PoisonedRAG for corpus poisoning attack.

📝 Citation and Reference

If you find this paper useful, please consider staring 🌟 this repo and citing 📑 our paper:

@article{zhou2025trustrag,
  title={Trustrag: Enhancing robustness and trustworthiness in rag},
  author={Zhou, Huichi and Lee, Kin-Hei and Zhan, Zhonghao and Chen, Yue and Li, Zhenhao and Wang, Zhaoyang and Haddadi, Hamed and Yilmaz, Emine},
  journal={arXiv preprint arXiv:2501.00879},
  year={2025}
}