app.py

import spaces
import os
import gradio as gr
from pdf2image import convert_from_path
from byaldi import RAGMultiModalModel
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import torchvision
import subprocess

def install_poppler():
    try:
        subprocess.run(["pdfinfo"], check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    except FileNotFoundError:
        print("Poppler not found. Installing...")
        subprocess.run("apt-get update", shell=True)
        subprocess.run("apt-get install -y poppler-utils", shell=True)

install_poppler()
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

RAG = RAGMultiModalModel.from_pretrained("vidore/colpali")
model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-2B-Instruct",
                                                        trust_remote_code=True, torch_dtype=torch.bfloat16).cuda().eval()
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", trust_remote_code=True)

@spaces.GPU()
def process_pdf_and_query(pdf_file, user_query):
    images = convert_from_path(pdf_file.name)
    num_images = len(images)
    RAG.index(
        input_path=pdf_file.name,
        index_name="image_index",
        store_collection_with_index=False,
        overwrite=True
    )
    results = RAG.search(user_query, k=1)
    if not results:
        return "No results found.", num_images
    image_index = results[0]["page_num"] - 1
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": images[image_index],
                },
                {"type": "text", "text": user_query},
            ],
        }
    ]
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    image_inputs, video_inputs = process_vision_info(messages)
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )
    inputs = inputs.to("cuda")
    generated_ids = model.generate(**inputs, max_new_tokens=50)
    generated_ids_trimmed = [
        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]
    output_text = processor.batch_decode(
        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )
    return output_text[0], num_images

footer = """
<div style="text-align: center; margin-top: 20px;">
    <a href="https://www.linkedin.com/in/pejman-ebrahimi-4a60151a7/" target="_blank">LinkedIn</a> |
    <a href="https://github.com/arad1367" target="_blank">GitHub</a> |
    <a href="https://arad1367.pythonanywhere.com/" target="_blank">Live demo of my PhD defense</a> |
    <a href="https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct" target="_blank">Qwen/Qwen2-VL-2B-Instruct</a> |
    <a href="https://github.com/AnswerDotAI/byaldi" target="_blank">Byaldi</a> |
    <a href="https://github.com/illuin-tech/colpali" target="_blank">ColPali</a>
    <br>
    Made with 💖 by Pejman Ebrahimi
</div>
"""

pdf_input = gr.File(label="Upload PDF")
query_input = gr.Textbox(label="Enter your query", placeholder="Ask a question about the PDF")
output_text = gr.Textbox(label="Model Answer")
output_images = gr.Textbox(label="Number of Images in PDF")
duplicate_button = gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button", color="green")

explanation_text = """
<div style='text-align: center; font-size: 16px;'>
    <h2 style='font-weight: bold;'>Multimodal RAG Overview</h2>
    <p>This application utilizes a Multimodal RAG (Retrieve-and-Generate) approach, enabling users to query information from PDF documents 
    by extracting relevant text and images. The ColPali model serves as a multimodal retriever, while the Byaldi library simplifies 
    the integration of ColPali. The Qwen/Qwen2-VL-2B-Instruct LLM enhances the generation of responses based on the retrieved content.</p>
</div>
"""

demo = gr.Interface(
    fn=process_pdf_and_query, 
    inputs=[pdf_input, query_input], 
    outputs=[output_text, output_images], 
    title="<div style='text-align: center; font-size: 24px; font-weight: bold;'>Multimodal RAG with Image Query</div>",
    description=explanation_text,
    theme='freddyaboulton/dracula_revamped'
)

demo.launch(debug=True)
demo.append(duplicate_button)
demo.append(gr.HTML(footer))