Undoc: RAFT 및 RAG 기반 의료상담 챗봇 모델 연구

A medical consultation chatbot using RAG and RAFT techniques on the Gemma 2 model, achieving high performance in the healthcare domain.

Demo

• Demo site: https://undoc.vercel.app
• Demo page Repo: https://github.com/jagaldol/undoc

The repository for the Demo page implementation can be found here.

Model

The RAFT-finetuned Gemma-2-2b-it model is optimized for healthcare question-answering tasks using the RAG.

Huggingface Model Card

Model Performance Comparison

Model	Median
Gemma-2-2b-it	0.80
Gemma-2-2b-it+RAG	0.89
DSF unit8	0.88
DSF unit8 + RAG	0.93
DSF	0.90
DSF + RAG	0.94
RAFT unit8	0.86
RAFT unit8 + RAG	0.93
RAFT	0.88
RAFT + RAG	0.96

We achieved a 16% performance increase from the base model!

we use RAGAS Semantic Similarity to measure performance.

How to Start

Follow the steps below to set up your development environment and get started with the project:

Set Up Environment Variables

Copy the .env.example file to .env and update the environment variables as needed:

$ cp .env.example .env

Open the .env file and update the variables with your specific configuration:

HF_TOKEN=your huggingface token
PINECONE_API_KEY=your pinecone api key
PINECONE_INDEX=your pinecone index name

Install Dependencies with Pipenv

Use pipenv to install the required dependencies and set up the virtual environment:

$ pipenv sync

This command will create a virtual environment if one does not already exist and install all dependencies as specified in the Pipfile.

Activate the Virtual Environment

To activate the virtual environment, run:

$ pipenv shell

Run the Application

After setting up the environment, you can start the application with the following command:

$ python main.py

Once the server is running, you can access the API at http://localhost:8000.

Inference

If you want a single response, run the inference.py script with a query to get a direct answer from the model:

$ python inference.py "배탈 난거 같은데 어떻게 해?"

This script takes a user query as input and generates a relevant response based on the AI model.

Structures

health-care-advisor
├── main.py
├── models
│   └── model_loader.py
├── notebooks
├── ragchain
│   ├── pipeline_setup.py
│   └── rag_chain.py
├── retriever
│   ├── hybrid_search_retriever.py
│   └── retriever_setup.py
├── template
│   ├── generation_prompt.py
│   └── script.py
└── utils
    └── environment.py

System Structure

Citation

• 초거대 AI 헬스케어 질의응답 데이터: AI 허브, 초거대 AI 헬스케어 질의응답 데이터
• Gemma 2 Model: "Gemma 2: Improving Open Language Models at a Practical Size", 2023.
• RAFT Methodology: "Adapting Language Model to Domain Specific RAG", arXiv preprint arXiv:2403.10131, 2023.
• RAGAS Evaluation: "RAGAS: Automated Evaluation of Retrieval Augmented Generation", 2023.

Collaborators

임영윤	안혜준