app.py

import gradio as gr
from doctr.models import ocr_predictor
import numpy as np

model = ocr_predictor(det_arch='db_resnet50', reco_arch='crnn_vgg16_bn', pretrained=True)

def extract_word_ready_table(image):
    if image is None:
        return "Please upload an image."
    
    img_array = np.array(image)
    result = model([img_array])
    json_export = result.export()
    
    markdown_rows = []
    # Boundaries for Name | Code | Statement | Group | Sub-Group | Normally
    col_bounds = [0.28, 0.35, 0.48, 0.62, 0.88] 

    for page in json_export['pages']:
        words_list = []
        for block in page['blocks']:
            for line in block['lines']:
                for word in line['words']:
                    y_top = word['geometry'][0][1]
                    y_bot = word['geometry'][1][1]
                    x_mid = (word['geometry'][0][0] + word['geometry'][1][0]) / 2
                    words_list.append({
                        'text': word['value'], 
                        'y_top': y_top, 
                        'y_bot': y_bot, 
                        'y_mid': (y_top + y_bot) / 2,
                        'x_mid': x_mid
                    })
        
        if not words_list: continue
        words_list.sort(key=lambda w: w['y_mid'])
        
        # 1. Smarter Row Grouping: We use a larger threshold (0.02) 
        # to catch text that is slightly above or below the main line
        rows = []
        current_row = [words_list[0]]
        for i in range(1, len(words_list)):
            # If word overlaps vertically with the current row, it's the SAME row
            if words_list[i]['y_top'] < current_row[-1]['y_bot'] + 0.01:
                current_row.append(words_list[i])
            else:
                rows.append(current_row)
                current_row = [words_list[i]]
        rows.append(current_row)
        
        # 2. Build the line
        for row in rows:
            slots = ["", "", "", "", "", ""]
            for w in row:
                x = w['x_mid']
                t = w['text']
                
                if x < col_bounds[0]: slots[0] += t + " "
                elif x < col_bounds[1]: slots[1] += t + " "
                elif x < col_bounds[2]: slots[2] += t + " "
                elif x < col_bounds[3]: slots[3] += t + " "
                elif x < col_bounds[4]: slots[4] += t + " "
                else: slots[5] += t + " "
            
            clean_slots = [s.strip() for s in slots]
            if any(clean_slots):
                # We use the Pipe (|) as the only separator
                markdown_rows.append("| " + " | ".join(clean_slots) + " |")

    return "\n".join(markdown_rows)

with gr.Blocks() as demo:
    gr.Markdown("## 📑 Word-Ready Accountancy Extractor")
    gr.Markdown("Forces wrapped text into a single line to prevent Word from merging cells incorrectly.")
    
    with gr.Row():
        with gr.Column():
            img_in = gr.Image(type="pil")
            btn = gr.Button("Extract for Word", variant="primary")
        with gr.Column():
            out = gr.Textbox(label="Result (One Line Per Row)", lines=25, elem_id="out_box")
            copy_btn = gr.Button("📋 Copy Table")
            copy_btn.click(None, None, None, js="""
                () => {
                    const text = document.querySelector('#output-text textarea').value;
                    navigator.clipboard.writeText(text);
                    alert('Copied! Now use Insert > Table > Convert Text to Table in Word.');
                }
            """)

    btn.click(extract_word_ready_table, inputs=img_in, outputs=out)

demo.launch()