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

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

# Call the Poppler installation check
install_poppler()

# Install flash-attn if not already installed
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

# Load the RAG Model and the Qwen2-VL-2B-Instruct model
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):
    # Convert the PDF to images
    images = convert_from_path(pdf_file.name)  
    num_images = len(images)

    # Indexing the PDF in RAG
    RAG.index(
        input_path=pdf_file.name,
        index_name="image_index",  # index will be saved at index_root/index_name/
        store_collection_with_index=False,
        overwrite=True
    )

    # Search the query in the RAG model
    results = RAG.search(user_query, k=1)
    if not results:
        return "No results found.", num_images

    # Retrieve the page number and process image
    image_index = results[0]["page_num"] - 1
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": images[image_index],
                },
                {"type": "text", "text": user_query},
            ],
        }
    ]

    # Generate text with the Qwen model
    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")
    
    # Generate the output response
    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


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")  

# CSS styling
css = """
body {
    background-color: #282a36;
    font-family: Arial, sans-serif;
    color: #f8f8f2;
}
h1 {
    text-align: center;
    font-size: 2.5em;
    font-weight: bold;
    margin-bottom: 20px;
}
footer {
    margin-top: 20px;
}
.duplicate-button {
    text-align: center;
    background-color: #50fa7b;
    color: #282a36;
    font-weight: bold;
    border: none;
    padding: 10px;
    cursor: pointer;
}
"""

description = """
### About Multimodal RAG
Multimodal Retrieval-Augmented Generation (RAG) integrates both images and text to provide more comprehensive and contextually accurate responses to user queries. It uses a retriever model like **ColPali** to search and retrieve relevant data and a large language model (LLM) like **Qwen/Qwen2-VL-2B-Instruct** to generate natural language answers based on the input.

In this demo, **ColPali** is used as a multimodal retriever, and the **Byaldi** library from answer.ai simplifies the use of ColPali. We are utilizing **Qwen2-VL-2B-Instruct** for text generation, enabling both text and image-based queries.
"""

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 <a href="https://github.com/arad1367" target="_blank">Pejman Ebrahimi</a>
</div>
"""

# Gradio Interface 
with gr.Blocks(theme='freddyaboulton/dracula_revamped', css=css) as demo:
    gr.Markdown("<h1>Multimodal RAG with Image Query</h1>")
    gr.Markdown(description)
    with gr.Row():
        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")
    
    gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button")
    gr.HTML(footer)

    demo.launch(debug=True)