Function calculate_offset

mainprogress/pdf_metadata_extractor.py:561–710

自动计算正文偏移量。 逻辑优化： 1. 先随机取 5 页。 2. 统计众数，若众数数量 < 4，则再随机取 5 页（不重复），共 10 页一起统计。 3. 将所有过程的图片、原始响应、解析结果保存到 initial_data/offset_log.json。

(pdf_path: str, client: AsyncOpenAI, model: str, initial_data_dir: str)

    2. 统计众数，若众数数量 < 4，则再随机取 5 页（不重复），共 10 页一起统计。
    3. 将所有过程的图片、原始响应、解析结果保存到 initial_data/offset_log.json。
    """
    log_entries = []
    try:
        doc = fitz.open(pdf_path)
        total_pages = len(doc)
        start_idx = int(total_pages * 0.2)
        end_idx = int(total_pages * 0.8)
        if end_idx <= start_idx:
            end_idx = total_pages - 1
            start_idx = 0

        pool = list(range(start_idx, end_idx + 1))
        if len(pool) < 5:
            doc.close()
            return None

        # 第一轮：取 5 页
        selected_pages_1 = random.sample(pool, 5)
        remaining_pool = [p for p in pool if p not in selected_pages_1]
        
        all_selected_indices = selected_pages_1[:]
        
        # 如果需要第二轮
        need_second_round = False
        
        # 临时存储第一轮结果用于判断
        first_round_results = []
        
        # 定义目标长边像素
        TARGET_LONG_EDGE = 1500

        for p in selected_pages_1:
            page = doc[p]
            rect = page.rect
            max_dim = max(rect.width, rect.height)
            
            # 修正逻辑：始终计算缩放比例
            if max_dim == 0:
                continue
            zoom = TARGET_LONG_EDGE / max_dim
            
            mat = fitz.Matrix(zoom, zoom)
            pix = page.get_pixmap(matrix=mat)
            img_data = pix.tobytes("jpeg")
            base64_image = base64.b64encode(img_data).decode('utf-8')
            
            raw_res = await fetch_single_offset(client, model, p + 1, base64_image)
            
            # 记录日志数据
            entry = {
                "physical_page": p + 1,
                "raw_response": raw_res,
                "parsed_offset": None,
                "image_base64_preview": base64_image[:100] + "..." # 仅存预览，避免 JSON 过大，实际图片可单独存如需
            }
            
            if raw_res.isdigit() or (raw_res.startswith('-') and raw_res[1:].isdigit()):
                val = int(raw_res)