Retrieval-Augmented Generation (RAG) Project

Overview

This project demonstrates a scalable Retrieval-Augmented Generation (RAG) system designed for medical question answering using a combination of a vector database and a local Large Language Model (LLM). It showcases modern machine learning and software engineering practices with the potential to integrate cloud-based solutions such as Azure Cognitive Search for production-ready applications.

The system retrieves relevant research findings from the PubMedQA dataset, embeds the content using Sentence Transformers, and stores them in Qdrant, a high-performance vector database. Queries are answered using a local LLM (e.g., Llama), augmented with the retrieved context.

---

Key Features

• Context-Enhanced Responses: Combines vector-based search with LLMs for accurate and context-aware answers.
• Modular Design: Supports local deployment with Qdrant and extensibility for cloud integration with Azure Cognitive Search.
• FastAPI Framework: Provides an intuitive and scalable API interface.
• Dockerized Environment: Simplifies deployment with separate configurations for development and production.

---

Architecture

1. Data Preparation:
   • The PubMedQA dataset is loaded, and relevant entries are embedded using Sentence Transformers (all-MiniLM-L6-v2).
2. Vector Storage:
   • Qdrant stores embeddings for efficient vector-based retrieval.
3. Query Pipeline:
   • User queries are vectorized and matched with relevant embeddings in Qdrant.
   • Retrieved context is passed to the local LLM (e.g., Llama) for an enriched response.
4. API Interaction:
   • Exposes endpoints for submitting queries (/ask) via FastAPI.

---

Technologies Used

• Language & Framework:
  • Python 3.10
  • FastAPI
• Machine Learning:
  • Hugging Face Datasets (PubMedQA)
  • Sentence Transformers
• Vector Database:
  • Qdrant
• Local LLM:
  • Llama (via LlamaFile)
• DevOps & Deployment:
  • Docker, Docker Compose
  • Multi-stage Dockerfile for optimized builds
• Cloud (Optional):
  • Azure Cognitive Search (ready for integration)

---

Setup Instructions

1. Prerequisites

• Docker and Docker Compose (v2+)
• Python 3.10+ (optional, for local testing)
• GPU (optional, for accelerated embeddings)

2. Clone the Repository

git clone https://github.com/tanle8/MediRAG.git
cd MediRAG

3. Install Dependencies

For Local Testing:

pip install -r requirements.txt

For Docker Deployment:

Build the Docker image with all dependencies:

docker compose up --build

---

Running the Project

Development Mode

Use the development configuration to enable live code reloading:

make dev-build

Access the FastAPI interface at http://localhost:80/docs.

Production Mode

For optimized builds and deployment:

make prod-build

---

Usage

API Endpoints

1. GET /:
   • Redirects to the Swagger UI at /docs.
2. POST /ask:
   • Submit a query and receive an enriched response.

Example Query:

curl -X POST "http://localhost:80/ask" -H "Content-Type: application/json" -d '{"query": "What are the long-term outcomes of laparoscopic surgery for hiatal hernia repair?"}'

Example Response:

{
  "response": "Laparoscopic surgery for hiatal hernia repair has shown positive long-term outcomes, with reduced recurrence rates compared to traditional open surgeries."
}

---

Environment Variables

Configure via .env:

LLM_TYPE=local
LOCAL_LLM_SERVICE_URL=http://host.docker.internal:8080/
AZURE_API_KEY=your-api-key  # For optional Azure integration

---

Folder Structure

.
├── webapp
│   ├── main.py                # FastAPI app
│   ├── vectorstore.py         # Qdrant operations
│   ├── embeddings.py          # Embedding generation
│   ├── llm.py                 # Interaction with local LLM
│   ├── __init__.py
├── requirements.txt           # Python dependencies
├── Dockerfile                 # Multi-stage Docker build
├── docker-compose.dev.yml     # Docker Compose for development
├── docker-compose.prod.yml    # Docker Compose for production
├── Makefile                   # Automation for builds and deployment
└── README.md                  # Project documentation

---

Future Enhancements

1. Cloud Integration:
   • Azure Cognitive Search for scalable vector-based retrieval.
   • Azure OpenAI for hosted GPT models.
2. Advanced Indexing:
   • Experiment with hybrid search (e.g., dense + sparse retrieval).
3. Scalability:
   • Kubernetes support for deploying across clusters.
4. Improved UI:
   • Build a web-based frontend for user-friendly interactions.

---

Contributing

We welcome contributions from the community! Please follow these steps:

1. Fork the repository.
2. Create a new branch (feature-xyz).
3. Commit your changes.
4. Open a pull request.

---

Acknowledgments

• Hugging Face for the PubMedQA dataset.
• Qdrant for the vector database.
• Sentence Transformers for efficient embedding generation.
• Uvicorn & FastAPI for the API framework.

---

License

This project is licensed under the MIT License. See LICENSE for more details.

---

Contact

For questions, feedback, or collaborations, please reach out:

• Email: [email protected]
• LinkedIn: Tan (David) LE
• GitHub: @tanle8