ocr_engine.py

import pytesseract
from pytesseract import Output
from pdf2image import convert_from_path
from PIL import Image
import os
import logging
import numpy as np

logger = logging.getLogger("ocr_engine")

def extract_text_and_conf(file_path: str) -> tuple[str, float]:
    """
    Extracts text AND confidence score from a PDF or Image.
    Returns: (text_content, average_confidence_0_to_100)
    """
    if not os.path.exists(file_path):
        return "", 0.0

    text_content = ""
    confidences = []
    
    try:
        images = []
        # 1. Load Images
        if file_path.lower().endswith('.pdf'):
            try:
                images = convert_from_path(file_path)
            except Exception as e:
                logger.error(f"PDF Convert Error: {e}")
                return "", 0.0
        elif file_path.lower().endswith(('.png', '.jpg', '.jpeg', '.tiff', '.bmp')):
            try:
                images = [Image.open(file_path)]
            except Exception as e:
                logger.error(f"Image Open Error: {e}")
                return "", 0.0
        
        # 2. Process Each Page
        for i, image in enumerate(images):
            # A. Get Layout-Preserved Text (Best for LLM)
            page_text = pytesseract.image_to_string(image)
            text_content += f"--- Page {i+1} ---\n{page_text}\n"
            
            # B. Get Confidence Data (Best for KPIs)
            # data_dict keys: ['level', 'page_num', 'block_num', 'par_num', 'line_num', 'word_num', 'left', 'top', 'width', 'height', 'conf', 'text']
            data = pytesseract.image_to_data(image, output_type=Output.DICT)
            
            # Filter valid confidences (ignore -1 which usually means whitespace/block info)
            for conf in data['conf']:
                # Tesseract returns -1 for structural elements (not words)
                if conf != -1:
                    confidences.append(conf)

        # 3. Calculate Average Confidence
        avg_conf = 0.0
        if confidences:
            avg_conf = sum(confidences) / len(confidences)
            
        return text_content.strip(), round(avg_conf, 2)

    except Exception as e:
        logger.error(f"OCR Critical Error: {e}")
        return "", 0.0