app.py

import gradio as gr
from PIL import Image
import json
from byaldi import RAGMultiModalModel
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import re

# Load models
def load_models():
    RAG = RAGMultiModalModel.from_pretrained("vidore/colpali")
    model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", trust_remote_code=True, torch_dtype=torch.float32)  # float32 for CPU
    processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", trust_remote_code=True)
    return RAG, model, processor

RAG, model, processor = load_models()

# Function for OCR
def extract_text_from_image(image):
    text_query = "Extract all the text in Sanskrit and English from the image."
    
    # Prepare message for Qwen model
    messages = [ 
        { 
            "role": "user", 
            "content": [
                {"type": "image", "image": image}, 
                {"type": "text", "text": text_query}
            ] 
        } 
    ]
    
    # Process the image
    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"
    ).to("cpu")  # Use CPU
    
    # Generate text
    with torch.no_grad():
        generated_ids = model.generate(**inputs, max_new_tokens=2000)
        generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        extracted_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    
    return extracted_text

# Function for keyword search
def search_keyword_in_text(extracted_text, keyword):
    keyword_lower = keyword.lower()
    sentences = extracted_text.split('. ')
    matched_sentences = []
    
    for sentence in sentences:
        if keyword_lower in sentence.lower():
            highlighted_sentence = re.sub(f'({re.escape(keyword)})', r'<mark>\1</mark>', sentence, flags=re.IGNORECASE)
            matched_sentences.append(highlighted_sentence)
    
    return matched_sentences if matched_sentences else ["No matches found."]

# Gradio App
def app_extract_text(image):
    extracted_text = extract_text_from_image(image)
    return extracted_text

def app_search_keyword(extracted_text, keyword):
    search_results = search_keyword_in_text(extracted_text, keyword)
    search_results_str = "<br>".join(search_results)
    return search_results_str


title_html = """
<h1><span class="gradient-text" id="text">IIT Roorkee - OCR and Document Search Web Application Prototype (ColPali implementation of the new Byaldi library + Huggingface transformers for Qwen2-VL.)</span></h1>
"""

# Gradio Interface
with gr.Blocks() as iface:

    gr.HTML(title_html)

    
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(type="pil", label="Upload an Image")
            extract_button = gr.Button("Extract Text")
            extracted_text_output = gr.Textbox(label="Extracted Text")
            
            extract_button.click(app_extract_text, inputs=image_input, outputs=extracted_text_output)

        with gr.Column():
            keyword_input = gr.Textbox(label="Enter keyword to search in extracted text", placeholder="Keyword")
            search_button = gr.Button("Search Keyword")
            search_results_output = gr.HTML(label="Search Results")

            search_button.click(app_search_keyword, inputs=[extracted_text_output, keyword_input], outputs=search_results_output)

# Launch Gradio App
iface.launch()