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):
    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
    )

    # 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},
            ],
        }
    ]


    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


css = """
    .duplicate-button {
        background-color: #6272a4;
        color: white;
        font-weight: bold;
        border-radius: 5px;
        margin-top: 20px;
        padding: 10px;
        text-align: center;
    }
    .gradio-container {
        background-color: #282a36;
        color: #f8f8f2;
        font-family: 'Courier New', Courier, monospace;
        padding: 20px;
        border-radius: 10px;
        box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1);
    }
"""

explanation = """
### Multimodal RAG with Image Query
This demo showcases the **Multimodal RAG (Retriever-Augmented Generation)** model. The RAG system integrates retrieval and generation, allowing it to retrieve relevant information from a multimodal database (like PDFs with text and images) and then generate detailed responses. 

We use **ColPali**, a state-of-the-art multimodal retriever, combined with the **Byaldi** library from **answer.ai**, which simplifies using ColPali. The language model used for generating answers is **Qwen/Qwen2-VL-2B-Instruct**, a powerful vision-language model capable of understanding both text and 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")

# Launch the Gradio app 
demo = gr.Interface(
    fn=process_pdf_and_query, 
    inputs=[pdf_input, query_input],  
    outputs=[output_text, output_images],  
    title="Multimodal RAG with Image Query - By Pejman Ebrahimi - Please like the space if it is useful",
    theme='freddyaboulton/dracula_revamped',
    css=css,
    description=explanation,
    allow_flagging="auto"
)

with demo:
    gr.HTML(footer)
    duplicate_button

demo.launch(debug=True)