app.py

import os
import gradio as gr
import asyncio
from langchain_core.prompts import PromptTemplate
from langchain_community.output_parsers.rail_parser import GuardrailsOutputParser
from langchain_community.document_loaders import PyPDFLoader
from langchain_google_genai import ChatGoogleGenerativeAI
import google.generativeai as genai
from langchain.chains.question_answering import load_qa_chain
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Gemini initialization and PDF QA function
async def initialize_gemini(file_path, question):
    genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))
    model = ChatGoogleGenerativeAI(model="gemini-pro", temperature=0.3)
    prompt_template = """Answer the question as precise as possible using the provided context. If the answer is
                          not contained in the context, say "answer not available in context" \n\n
                          Context: \n {context}?\n
                          Question: \n {question} \n
                          Answer:
                        """
    prompt = PromptTemplate(template=prompt_template, input_variables=["context", "question"])
    if os.path.exists(file_path):
        pdf_loader = PyPDFLoader(file_path)
        pages = pdf_loader.load_and_split()
        context = "\n".join(str(page.page_content) for page in pages[:100])
        stuff_chain = load_qa_chain(model, chain_type="stuff", prompt=prompt)
        stuff_answer = stuff_chain({"input_documents": pages, "question": question, "context": context}, return_only_outputs=True)
        return stuff_answer['output_text']
    else:
        return "Error: Unable to process the document. Please ensure the PDF file is valid."

# Mistral model initialization
def initialize_mistral():
    model_path = "nvidia/Mistral-NeMo-Minitron-8B-Base"
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    dtype = torch.bfloat16
    model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=dtype)  # Removed device_map parameter
    return tokenizer, model

# Mistral text generation function
def generate_mistral_text(prompt, tokenizer, model):
    inputs = tokenizer.encode(prompt, return_tensors='pt').to(model.device)
    outputs = model.generate(inputs, max_length=100)
    return tokenizer.decode(outputs[0])

# Initialize Mistral model
mistral_tokenizer, mistral_model = initialize_mistral()

# Gradio interface function
async def pdf_qa(file, question):
    gemini_answer = await initialize_gemini(file.name, question)
    mistral_prompt = f"Based on this answer: '{gemini_answer}', provide a brief summary:"
    mistral_summary = generate_mistral_text(mistral_prompt, mistral_tokenizer, mistral_model)
    return f"Gemini Answer:\n{gemini_answer}\n\nMistral Summary:\n{mistral_summary}"

# Define Gradio Interface
input_file = gr.File(label="Upload PDF File")
input_question = gr.Textbox(label="Ask about the document")
output_text = gr.Textbox(label="Answer and Summary")

# Create Gradio Interface
gr.Interface(
    fn=pdf_qa,
    inputs=[input_file, input_question],
    outputs=output_text,
    title="PDF Question Answering System with Gemini and Mistral",
    description="Upload a PDF file, ask questions about the content, and get answers from Gemini with a summary from Mistral."
).launch()