💊 Drug Information Chatbot

A Deep Learning chatbot powered by Flan-T5 + LoRA fine-tuning and Retrieval-Augmented Generation (RAG) for answering drug-related questions.

🌟 Features

• Fine-tuned Language Model: Flan-T5-Base with LoRA adapters trained on medical Q&A data
• RAG (Retrieval-Augmented Generation): FAISS-based vector search for accurate drug information
• Interactive UI: Built with Streamlit for easy interaction
• Source Attribution: Shows sources used for each answer
• Medical Database: Information from DailyMed and MedQuAD datasets

📋 Prerequisites

• Python 3.8 or higher
• CUDA-capable GPU (optional, but recommended for faster inference)
• Git (for cloning the repository)

🚀 Installation

1. Clone the Repository

git clone <repository-url>
cd deep-learning-final-project

2. Create a Virtual Environment (Recommended)

On Windows:

python -m venv venv
venv\Scripts\activate

On macOS/Linux:

python -m venv venv
source venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

Note: If you have a CUDA-enabled GPU and want to use GPU acceleration:

1. Install PyTorch with CUDA support from pytorch.org
2. Replace faiss-cpu with faiss-gpu in requirements.txt

📁 Project Structure

deep-learning-final-project/
├── app/
│   ├── streamlit_app.py          # Main Streamlit application
│   └── __init__.py
├── notebooks/
│   ├── faiss_rag_builder.py      # RAG retriever implementation
│   ├── finetuning_model.ipynb    # Model fine-tuning notebook
│   └── ...
├── data/
│   ├── faiss/                     # FAISS index and metadata
│   ├── finetuning/                # Training data
│   └── rag/                       # RAG knowledge base
├── models/
│   └── drug_qna_lora/
│       └── final/                 # Fine-tuned model checkpoint
├── requirements.txt               # Python dependencies
└── README.md                      # This file

▶️ Running the Application

Prerequisites Before Running

Make sure you have:

1. ✅ Trained model files in models/drug_qna_lora/final/
2. ✅ FAISS index files in data/faiss/
   • drug_knowledge.index
   • metadata.pkl
   • config.json

If these files are missing, you'll need to:

• Run the fine-tuning notebook: notebooks/finetuning_model.ipynb
• Build the RAG index: notebooks/faiss_rag_builder.ipynb

Start the Streamlit App

streamlit run app/streamlit_app.py

The application will automatically open in your default web browser at http://localhost:8501

Alternative: Specify Port

streamlit run app/streamlit_app.py --server.port 8080

🎯 Using the Chatbot

1. Ask Questions: Type your drug-related questions in the chat input
2. View Answers: The chatbot will provide answers based on the trained model
3. Check Sources: Expand the "Sources Used" section to see the retrieved documents
4. Try Examples: Click the example questions in the sidebar
5. Toggle RAG: Enable/disable RAG in the settings (sidebar)
6. Clear Chat: Use the "Clear Chat" button to start a new conversation

Example Questions

• "What is the dosage of Amoxicillin?"
• "What are the warnings and precautions of Atorvastatin?"
• "How does Albuterol work?"
• "What are the side effects of Ibuprofen?"
• "When should I not take Amoxicillin?"

⚙️ Configuration

Model Settings

Edit the paths in app/streamlit_app.py:

MODEL_DIR = ROOT / "models" / "drug_qna_lora" / "final"
FAISS_DIR = ROOT / "data" / "faiss"

Generation Parameters

Modify the generation parameters in the generate_answer() function:

outputs = model.generate(
    **inputs,
    max_length=512,        # Maximum output length
    num_beams=4,           # Beam search width
    repetition_penalty=1.2, # Avoid repetition
    no_repeat_ngram_size=3, # N-gram blocking
    early_stopping=True,
)

🛠️ Troubleshooting

Issue: Model Not Loading

Error: "❌ Model not loaded"

Solution:

• Ensure you have trained the model first
• Check that the model files exist in models/drug_qna_lora/final/
• Run the fine-tuning notebook to create the model

Issue: RAG Not Available

Error: "⚠️ RAG not available"

Solution:

• Check that FAISS index files exist in data/faiss/
• Run the faiss_rag_builder.ipynb notebook to build the index
• Verify the faiss-cpu package is installed

Issue: Out of Memory

Solution:

• Reduce batch size or max_length in generation
• Use CPU instead of GPU (set device_map="cpu")
• Close other memory-intensive applications

Issue: Slow Performance

Solution:

• Use GPU if available (install CUDA-enabled PyTorch)
• Reduce num_beams parameter
• Cache the model loading (already implemented with @st.cache_resource)

📊 Training Your Own Model

To train the model from scratch:

1. Prepare your dataset in data/finetuning/
2. Open and run notebooks/finetuning_model.ipynb
3. The trained model will be saved to models/drug_qna_lora/final/

🗂️ Building the RAG Index

To rebuild the FAISS index:

1. Ensure your knowledge base is in data/rag/knowledge_base.json
2. Open and run notebooks/faiss_rag_builder.ipynb
3. The index will be saved to data/faiss/

Important Disclaimer

This chatbot is for educational purposes only.

• Do NOT use this as a substitute for professional medical advice
• Always consult with qualified healthcare professionals for medical decisions
• The information provided may not be complete or up-to-date
• This is a student project for a Deep Learning course

Project Information

• Course: Deep Learning Final Project
• Technology Stack:
  • Flan-T5-Base (Google)
  • LoRA (Low-Rank Adaptation)
  • FAISS (Facebook AI Similarity Search)
  • Sentence Transformers
  • Streamlit
• Data Sources:
  • DailyMed (drug information)
  • MedQuAD (medical Q&A dataset)

📝 License

This project is for educational purposes. Please check the licenses of the underlying models and datasets:

• Flan-T5 License
• DailyMed Terms