Create app.py

app.py

@@ -0,0 +1,98 @@
+import os
+import asyncio
+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
+from functools import lru_cache
+import concurrent.futures
+import pymupdf
+
+# Configure Gemini API
+genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))
+
+# Load Mistral model (lazy loading)
+model_path = "nvidia/Mistral-NeMo-Minitron-8B-Base"
+mistral_tokenizer = None
+mistral_model = None
+
+def load_mistral_model():
+    global mistral_tokenizer, mistral_model
+    if mistral_tokenizer is None or mistral_model is None:
+        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)
+
+@lru_cache(maxsize=100)
+def get_pdf_context(file_path):
+    doc = pymupdf.open(file_path)
+    text = ""
+    for page in doc:
+        text += page.get_text()
+    return text[:10000]  # Limit context to first 10000 characters
+
+async 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"])
+    
+    context = get_pdf_context(file_path)
+    stuff_chain = load_qa_chain(model, chain_type="stuff", prompt=prompt)
+    stuff_answer = await stuff_chain.arun({"input_documents": [context], "question": question, "context": context})
+    return stuff_answer
+
+async def process_image(image, question):
+    model = genai.GenerativeModel('gemini-pro-vision')
+    response = await model.generate_content_async([image, question])
+    return response.text
+
+async def generate_mistral_followup(answer):
+    load_mistral_model()
+    mistral_prompt = f"Based on this answer: {answer}\nGenerate a follow-up question:"
+    mistral_inputs = mistral_tokenizer.encode(mistral_prompt, return_tensors='pt').to(mistral_model.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)
+    return mistral_output
+
+async def process_input(file, image, question):
+    try:
+        if file is not None:
+            gemini_answer = await process_pdf(file.name, question)
+        elif image is not None:
+            gemini_answer = await process_image(image, question)
+        else:
+            return "Please upload a PDF file or an image."
+
+        mistral_followup = await 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)}"
+
+# Gradio Interface
+with gr.Blocks() as demo:
+    gr.Markdown("# Optimized 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=lambda file, image, question: asyncio.run(process_input(file, image, question)), 
+                        inputs=[input_file, input_image, input_question], 
+                        outputs=output_text)
+
+if __name__ == "__main__":
+    demo.launch()