This repository contains code for our AMIA'2024 paper "MKRAG: Medical Knowledge Retrieval Augmented Generation for Medical Question Answering".

The system combines:

• Large Language Models (Vicuna-7B) for answer generation
• Knowledge Graph integration for medical facts
• Contriever for relevant fact retrieval

pip install -r requirements.txt

medicalQA/
├── Release/
│   ├── main.py          # Main execution script
│   ├── generator.py     # Model generation wrapper
│   └── utils.py         # Utility functions
├── data/
│   ├── ddb/            # DrugBank database files
│   │   ├── ddb_names.json
│   │   └── ddb_relas.json
│   └── [Other data directories...]
└── [Other project directories...]

1. Download required model weights:

   • Vicuna-7B model: We used Vicuna-7B-v1 in our research. Due to licensing restrictions, we cannot publicly distribute the model. If you would like to reproduce our results, please contact [us](yucheng.shi (AT) uga (DOT) edu) via email for guidance and access instructions.
   • SapBERT/Contriever model weights are automatically downloaded from Huggingface.

2. Prepare knowledge graph data:

   • Place DrugBank database files in data/ddb/
   • You can find more information here.

python main.py --model vicuna --device cuda 

• --model: Model type ('vicuna', 'gpt2', etc.)
• --device: Computing device ('cuda' or 'cpu')
• --path: Path to medical QA dataset
• --fact_number: Number of facts to retrieve (default: 8)
• --loademb: Whether to load pre-computed embeddings. You can download it here.

• Uses Contriever for retrieving relevant medical facts
• Integrates DrugBank knowledge graph

• Employs Vicuna-7B for generating answers
• Incorporates retrieved knowledge into prompts

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

@article{shi2023mededit,
  title={Mededit: Model editing for medical question answering with external knowledge bases},
  author={Shi, Yucheng and Xu, Shaochen and Liu, Zhengliang and Liu, Tianming and Li, Xiang and Liu, Ninghao},
  journal={arXiv preprint arXiv:2309.16035},
  year={2023}
}