app.py

import os
import gradio as gr
from langchain_core.prompts import PromptTemplate
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
from PIL import Image
import io

# Configure Gemini API
genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))

# Load Mistral model
model_path = "nvidia/Mistral-NeMo-Minitron-8B-Base"
mistral_tokenizer = AutoTokenizer.from_pretrained(model_path)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
dtype = torch.bfloat16
mistral_model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=dtype, device_map=device)

def process_pdf(file_path, question):
    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"])
    
    pdf_loader = PyPDFLoader(file_path)
    pages = pdf_loader.load_and_split()
    context = "\n".join(str(page.page_content) for page in pages[:200])
    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']

def process_image(image, question):
    model = genai.GenerativeModel('gemini-pro-vision')
    response = model.generate_content([image, question])
    return response.text

def generate_mistral_followup(answer):
    mistral_prompt = f"Based on this answer: {answer}\nGenerate a follow-up question:"
    mistral_inputs = mistral_tokenizer.encode(mistral_prompt, return_tensors='pt').to(device)
    with torch.no_grad():
        mistral_outputs = mistral_model.generate(mistral_inputs, max_length=200)
    mistral_output = mistral_tokenizer.decode(mistral_outputs[0], skip_special_tokens=True)
    return mistral_output

def process_input(file, image, question):
    try:
        if file is not None:
            gemini_answer = process_pdf(file.name, question)
        elif image is not None:
            gemini_answer = process_image(image, question)
        else:
            return "Please upload a PDF file or an image."

        mistral_followup = generate_mistral_followup(gemini_answer)
        combined_output = f"Gemini Answer: {gemini_answer}\n\nMistral Follow-up: {mistral_followup}"
        return combined_output
    except Exception as e:
        return f"An error occurred: {str(e)}"

# Define Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# Multi-modal RAG Knowledge Retrieval using Gemini API and Mistral Model")
    
    with gr.Row():
        with gr.Column():
            input_file = gr.File(label="Upload PDF File")
            input_image = gr.Image(type="pil", label="Upload Image")
        input_question = gr.Textbox(label="Ask about the document or image")
    
    output_text = gr.Textbox(label="Answer - Combined Gemini and Mistral")
    
    submit_button = gr.Button("Submit")
    submit_button.click(fn=process_input, inputs=[input_file, input_image, input_question], outputs=output_text)

demo.launch()