import os
import gradio as gr
import asyncio
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

# 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)

async def initialize(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"])
    
    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[:30])
        stuff_chain = load_qa_chain(model, chain_type="stuff", prompt=prompt)
        stuff_answer = await stuff_chain.acall({"input_documents": pages, "question": question, "context": context}, return_only_outputs=True)
        gemini_answer = stuff_answer['output_text']
        
        # Use Mistral model for additional text generation
        mistral_prompt = f"Based on this answer: {gemini_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=50)
        mistral_output = mistral_tokenizer.decode(mistral_outputs[0], skip_special_tokens=True)
        
        combined_output = f"Gemini Answer: {gemini_answer}\n\nMistral Follow-up: {mistral_output}"
        return combined_output
    else:
        return "Error: Unable to process the document. Please ensure the PDF file is valid."

# 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 - Combined Gemini and Mistral")

async def pdf_qa(file, question):
    answer = await initialize(file.name, question)
    return answer

# Create Gradio Interface
gr.Interface(
    fn=pdf_qa,
    inputs=[input_file, input_question],
    outputs=output_text,
    title="RAG Knowledge Retrieval using Gemini API and Mistral Model",
    description="Upload a PDF file and ask questions about the content."
).launch()