app.py

import os
import torch
import torch.backends.cudnn as cudnn
import streamlit as st
from langchain_community.document_loaders import PyPDFLoader
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain.prompts import PromptTemplate
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline

# Enable CUDA optimizations if available
if torch.cuda.is_available():
    cudnn.benchmark = True

# Step 1: Load the PDF and create a vector store
@st.cache_resource
def load_pdf_to_vectorstore(pdf_path):
    # Load and split PDF
    loader = PyPDFLoader(pdf_path)
    documents = loader.load()
    
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=1000,
        chunk_overlap=20,
        separators=["\n\n", "\n", ".", " ", ""]
    )
    
    chunks = text_splitter.split_documents(documents)
    
    # Create embeddings and vector store
    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-MiniLM-L6-v2"
    )
    vectorstore = FAISS.from_documents(chunks, embeddings)
    
    return vectorstore

# Step 2: Initialize the LaMini model
@st.cache_resource
def setup_model():
    model_id = "MBZUAI/LaMini-Flan-T5-248M"  # Using smaller model for faster inference
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForSeq2SeqLM.from_pretrained(
        model_id,
        # Removed low_cpu_mem_usage parameter
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
    )
    
    if torch.cuda.is_available():
        model = model.cuda()
    
    pipe = pipeline(
        "text2text-generation",
        model=model,
        tokenizer=tokenizer,
        max_length=256,
        do_sample=False,
        temperature=0.3,
        top_p=0.95,
        device=0 if torch.cuda.is_available() else -1,
        batch_size=1
    )
    return pipe

# Step 3: Generate a response using the model and vector store
def generate_response(pipe, vectorstore, user_input):
    # Get relevant context
    docs = vectorstore.similarity_search(user_input, k=2)
    context = "\n".join([
        f"Page {doc.metadata.get('page', 'unknown')}: {doc.page_content}"
        for doc in docs
    ])
    
    # Create prompt
    prompt = PromptTemplate(
        input_variables=["context", "question"],
        template="""
        Using the following medical text excerpts, answer the question.
        If the information isn't clearly provided in the context, or if you're unsure, please say so and recommend consulting a healthcare professional.
        
        Context: {context}
        
        Question: {question}
        
        Answer (citing relevant page numbers when possible):"""
    )
    
    # Generate response using the new method
    prompt_text = prompt.format(context=context, question=user_input)
    response = pipe(prompt_text)[0]['generated_text']
    
    return response

# Cache responses for repeated questions
@st.cache_data
def cached_generate_response(user_input, _pipe, _vectorstore):
    return generate_response(_pipe, _vectorstore, user_input)

# Batch processing for multiple questions
def batch_generate_responses(pipe, vectorstore, questions, batch_size=4):
    responses = []
    for i in range(0, len(questions), batch_size):
        batch = questions[i:i + batch_size]
        batch_responses = [generate_response(pipe, vectorstore, q) for q in batch]
        responses.extend(batch_responses)
    return responses

# Streamlit UI
def main():
    st.title("Medical Chatbot Assistant 🏥")
    
    # Use the PDF file from the root directory
    pdf_path = "Medical_book.pdf"
    
    if os.path.exists(pdf_path):
        # Initialize progress
        progress_text = "Operation in progress. Please wait."
        
        # Load vector store and model with progress indication
        with st.spinner("Loading PDF and initializing model..."):
            vectorstore = load_pdf_to_vectorstore(pdf_path)
            pipe = setup_model()
            st.success("Ready to answer questions!")
        
        # Create a chat-like interface
        if "messages" not in st.session_state:
            st.session_state.messages = []

        # Display chat history
        for message in st.session_state.messages:
            with st.chat_message(message["role"]):
                st.markdown(message["content"])

        # User input
        if prompt := st.chat_input("Ask your medical question:"):
            # Add user message to chat history
            st.session_state.messages.append({"role": "user", "content": prompt})
            with st.chat_message("user"):
                st.markdown(prompt)

            # Generate and display response
            with st.chat_message("assistant"):
                with st.spinner("Generating response..."):
                    response = cached_generate_response(prompt, pipe, vectorstore)
                    st.markdown(response)
                    # Add assistant message to chat history
                    st.session_state.messages.append({"role": "assistant", "content": response})
                    
    else:
        st.error("The file 'Medical_book.pdf' was not found in the root directory.")


main()