diff --git a/demo/demo.py b/demo/demo.py index 037b0a5d..3cb1de22 100644 --- a/demo/demo.py +++ b/demo/demo.py @@ -8,7 +8,7 @@ import httpx from mineru.cli_old import api_client as _api_client from mineru.cli_old.common import image_suffixes, office_suffixes, pdf_suffixes -from mineru.utils.guess_suffix_or_lang import guess_suffix_by_path +from mineru.parser.file_type import guess_suffix_by_path SUPPORTED_INPUT_SUFFIXES = set(pdf_suffixes + image_suffixes + office_suffixes) diff --git a/mineru/backend/__init__.py b/mineru/backend/__init__.py index 1e17167c..818730ae 100644 --- a/mineru/backend/__init__.py +++ b/mineru/backend/__init__.py @@ -1 +1,3 @@ # Copyright (c) Opendatalab. All rights reserved. + +__all__: list[str] = [] diff --git a/mineru/backend/analysis/__init__.py b/mineru/backend/analysis/__init__.py index 04cbaa22..38bccbb3 100644 --- a/mineru/backend/analysis/__init__.py +++ b/mineru/backend/analysis/__init__.py @@ -1,2 +1,4 @@ # Copyright (c) Opendatalab. All rights reserved. """文档 model-list 生产层。""" + +__all__: list[str] = [] diff --git a/mineru/backend/analysis/office.py b/mineru/backend/analysis/office.py index b8c490a0..fe23b644 100644 --- a/mineru/backend/analysis/office.py +++ b/mineru/backend/analysis/office.py @@ -6,7 +6,7 @@ from __future__ import annotations import time from io import BytesIO -from mineru.model.flash import DocModel, DocxModel, PptModel, PptxModel, XlsModel, XlsxModel +from ...model.flash import DocModel, DocxModel, PptModel, PptxModel, XlsModel, XlsxModel from .contracts import AnalysisResult, OfficeSuffix diff --git a/mineru/backend/analysis/pdf/__init__.py b/mineru/backend/analysis/pdf/__init__.py index 916f622e..925fa648 100644 --- a/mineru/backend/analysis/pdf/__init__.py +++ b/mineru/backend/analysis/pdf/__init__.py @@ -1,2 +1,4 @@ # Copyright (c) Opendatalab. All rights reserved. """PDF 分析流水线。""" + +__all__: list[str] = [] diff --git a/mineru/backend/analysis/pdf/constants.py b/mineru/backend/analysis/pdf/constants.py index 02c66b90..113f3ff5 100644 --- a/mineru/backend/analysis/pdf/constants.py +++ b/mineru/backend/analysis/pdf/constants.py @@ -5,7 +5,7 @@ from __future__ import annotations import re -from mineru.types import RAW_ALGORITHM, RAW_CAPTION, RAW_FOOTNOTE, RAW_FORMULA_NUMBER, RAW_PHONETIC, BlockType +from ....types import RAW_ALGORITHM, RAW_CAPTION, RAW_FOOTNOTE, RAW_FORMULA_NUMBER, RAW_PHONETIC, BlockType LAYOUT_BASE_BATCH_SIZE = 1 MFR_BASE_BATCH_SIZE = 16 diff --git a/mineru/backend/analysis/pdf/formulas.py b/mineru/backend/analysis/pdf/formulas.py index 64afa467..40dba8d0 100644 --- a/mineru/backend/analysis/pdf/formulas.py +++ b/mineru/backend/analysis/pdf/formulas.py @@ -10,10 +10,10 @@ import cv2 import numpy as np from PIL import Image -from mineru.backend.local_model_runtime import HybridLocalModelContext -from mineru.utils.bbox_utils import normalize_to_int_bbox -from mineru.utils.ocr_utils import get_rotate_crop_image_for_text_rec -from mineru.utils.text_utils import build_tagged_formula_content +from ....model.runtime.hybrid import HybridLocalModelContext +from ....utils.geometry import normalize_to_int_bbox +from ....model.ocr.image import get_rotate_crop_image_for_text_rec +from ....utils.text import build_tagged_formula_content from ....types import RAW_FORMULA_NUMBER, BlockType diff --git a/mineru/backend/analysis/pdf/geometry.py b/mineru/backend/analysis/pdf/geometry.py index 91134e27..e15a7296 100644 --- a/mineru/backend/analysis/pdf/geometry.py +++ b/mineru/backend/analysis/pdf/geometry.py @@ -10,8 +10,8 @@ import cv2 import numpy as np from loguru import logger -from mineru.types import BBox -from mineru.utils.bbox_utils import normalize_to_int_bbox +from ....types import BBox +from ....utils.geometry import normalize_to_int_bbox def _normalize_page_size(page_image: Any) -> tuple[int, int]: diff --git a/mineru/utils/pdf_image_tools.py b/mineru/backend/analysis/pdf/images.py similarity index 80% rename from mineru/utils/pdf_image_tools.py rename to mineru/backend/analysis/pdf/images.py index ed571612..2892c3c1 100644 --- a/mineru/utils/pdf_image_tools.py +++ b/mineru/backend/analysis/pdf/images.py @@ -8,21 +8,49 @@ import threading import time from concurrent.futures import ALL_COMPLETED, Future, ProcessPoolExecutor, wait from concurrent.futures.process import BrokenProcessPool -from io import BytesIO +import math from typing import Any, Callable, Literal import numpy as np import pypdfium2 as pdfium from loguru import logger -from PIL import Image, ImageOps +from PIL import Image + +from ....model.flash.pdf.pdfium import close_pdfium_child, pdfium_guard +from ....model.flash.pdf.raster import image_to_b64str, page_to_image +from ....types import BBox, IntBBox +from ....utils.geometry import normalize_to_int_bbox +from ....utils.platform import is_windows_environment + + +class ImageType: + """限定 PDF 页面渲染支持的两种图像返回形态。""" + + PIL = "pil_img" + BASE64 = "base64_img" + + +def _positive_int_env(name: str, default: int) -> int: + """读取正整数环境变量，缺失或非法时返回默认值。""" + raw_value = os.getenv(name) + if raw_value is None: + return default + try: + value = int(raw_value) + except ValueError: + return default + return value if value > 0 else default + + +def get_load_images_timeout() -> int: + """返回 PDF 批量渲染的超时秒数。""" + return _positive_int_env("MINERU_PDF_RENDER_TIMEOUT", 300) + + +def get_load_images_threads() -> int: + """返回 PDF 批量渲染允许使用的进程数。""" + return _positive_int_env("MINERU_PDF_RENDER_THREADS", 3) -from ..types import BBox -from .bbox_utils import normalize_to_int_bbox -from .check_sys_env import is_windows_environment -from .enum_class import ImageType -from .os_env_config import get_load_images_threads, get_load_images_timeout -from .pdf_reader import image_to_b64str, page_to_image -from .pdfium_guard import close_pdfium_child, pdfium_guard DEFAULT_PDF_IMAGE_DPI = 200 # DEFAULT_PDF_IMAGE_DPI = 144 @@ -34,6 +62,7 @@ PDF_RENDER_KILL_JOIN_TIMEOUT_SECONDS = 0.1 _pdf_render_executor: ProcessPoolExecutor | None = None _pdf_render_executor_lock = threading.Lock() +_pdf_render_atexit_registered = False _pdf_render_spawn_submit_lock = threading.Lock() _pdf_render_spawn_submit_executor_id: int | None = None _pdf_render_spawn_submit_count = 0 @@ -44,16 +73,7 @@ def pdf_page_to_image( dpi: int = DEFAULT_PDF_IMAGE_DPI, image_type: Literal["pil_img", "base64_img"] = ImageType.PIL, ) -> dict[str, Any]: - """Convert pdfium.PdfDocument to image, Then convert the image to base64. - - Args: - page (_type_): pdfium.PdfPage - dpi (int, optional): reset the dpi of dpi. Defaults to DEFAULT_PDF_IMAGE_DPI. - image_type (ImageType, optional): The type of image to return. Defaults to ImageType.PIL. - - Returns: - dict: {'img_base64': str, 'img_pil': pil_img, 'scale': float } - """ + """将单个 PDFium 页面渲染为 Pillow 图像或 base64 载荷。""" pil_img, scale = page_to_image(page, dpi=dpi) image_dict: dict[str, Any] = {"scale": scale} if image_type == ImageType.BASE64: @@ -109,6 +129,7 @@ def _close_image_dicts(images_list: list[dict[str, Any]] | None) -> None: def _calculate_render_process_count(total_pages: int, threads: int, cpu_count: int | None = None) -> int: + """按页数、配置和 CPU 数量计算实际渲染进程数。""" requested_threads = max(1, threads) available_cpus = max(1, cpu_count if cpu_count is not None else (os.cpu_count() or 1)) page_limited_threads = max(1, total_pages // MIN_PAGES_PER_RENDER_PROCESS) @@ -125,6 +146,7 @@ def _build_render_page_ranges( end_page_id: int, process_count: int, ) -> list[tuple[int, int]]: + """把闭区间页范围均匀拆分为指定数量的子范围。""" total_pages = end_page_id - start_page_id + 1 base_pages, remainder = divmod(total_pages, process_count) page_ranges = [] @@ -145,12 +167,14 @@ def _get_render_process_plan( threads: int, cpu_count: int | None = None, ) -> tuple[int, list[tuple[int, int]]]: + """同时返回实际进程数及其页面分片计划。""" total_pages = end_page_id - start_page_id + 1 actual_threads = _calculate_render_process_count(total_pages, threads, cpu_count) return actual_threads, _build_render_page_ranges(start_page_id, end_page_id, actual_threads) def _get_pdf_render_pool_capacity(cpu_count: int | None = None) -> int: + """返回持久 PDF 渲染进程池的最大容量。""" available_cpus = max(1, cpu_count if cpu_count is not None else (os.cpu_count() or 1)) configured_threads = max(1, get_load_images_threads()) return min( @@ -161,6 +185,7 @@ def _get_pdf_render_pool_capacity(cpu_count: int | None = None) -> int: def _exit_pdf_render_worker_when_parent_exits() -> None: + """等待父进程退出后立即终止孤立渲染 worker。""" parent = multiprocessing.parent_process() if parent is None: return @@ -169,6 +194,7 @@ def _exit_pdf_render_worker_when_parent_exits() -> None: def _install_pdf_render_parent_exit_watcher() -> None: + """在渲染 worker 中安装父进程退出监听线程。""" watcher = threading.Thread( target=_exit_pdf_render_worker_when_parent_exits, name="mineru-pdf-render-parent-exit-watcher", @@ -178,6 +204,7 @@ def _install_pdf_render_parent_exit_watcher() -> None: def _create_pdf_render_executor(max_workers: int) -> ProcessPoolExecutor: + """使用安全 multiprocessing 上下文创建 PDF 渲染进程池。""" if is_windows_environment(): return ProcessPoolExecutor(max_workers=max_workers, initializer=_install_pdf_render_parent_exit_watcher) @@ -234,6 +261,7 @@ def _submit_pdf_render_task( def _get_pdf_render_future_states( future_to_range: dict[Future[Any], tuple[int, int]], ) -> list[dict[str, Any]]: + """汇总每个渲染 Future 的页面范围和运行状态。""" states = [] for future, (range_start, range_end) in future_to_range.items(): try: @@ -257,6 +285,7 @@ def _get_pdf_render_future_states( def _get_pdf_render_worker_states(executor: ProcessPoolExecutor) -> list[dict[str, Any]]: + """读取进程池 worker 的 pid、存活状态和退出码。""" try: process_map = getattr(executor, "_processes", None) or {} processes = list(process_map.values()) @@ -288,10 +317,14 @@ def _get_pdf_render_worker_states(executor: ProcessPoolExecutor) -> list[dict[st def _get_pdf_render_executor() -> ProcessPoolExecutor: - global _pdf_render_executor + """惰性创建并复用持久 PDF 渲染进程池。""" + global _pdf_render_atexit_registered, _pdf_render_executor with _pdf_render_executor_lock: if _pdf_render_executor is None: + if not _pdf_render_atexit_registered: + atexit.register(shutdown_pdf_render_executor) + _pdf_render_atexit_registered = True max_workers = _get_pdf_render_pool_capacity() _pdf_render_executor = _create_pdf_render_executor(max_workers=max_workers) logger.debug(f"Created persistent PDF render executor with max_workers={max_workers}") @@ -303,6 +336,7 @@ def _recycle_pdf_render_executor( *, terminate_processes: bool, ) -> None: + """从全局缓存移除指定进程池并按需终止 worker。""" global _pdf_render_executor if executor is None: @@ -324,6 +358,7 @@ def _recycle_pdf_render_executor( def shutdown_pdf_render_executor() -> None: + """关闭当前持久 PDF 渲染进程池。""" global _pdf_render_executor with _pdf_render_executor_lock: @@ -337,9 +372,6 @@ def shutdown_pdf_render_executor() -> None: ) -atexit.register(shutdown_pdf_render_executor) - - def load_images_from_pdf_bytes_range( pdf_bytes: bytes, dpi: int = DEFAULT_PDF_IMAGE_DPI, @@ -349,6 +381,7 @@ def load_images_from_pdf_bytes_range( timeout: int | None = None, threads: int | None = None, ) -> list[dict[str, Any]]: + """在持久进程池中批量渲染指定闭区间页面。""" if end_page_id < start_page_id: return [] @@ -500,6 +533,7 @@ def load_images_from_pdf_core( end_page_id: int | None = None, image_type: Literal["pil_img", "base64_img"] = ImageType.PIL, ) -> list[dict[str, Any]]: + """在当前进程中依次渲染 PDF 页面范围。""" images_list = [] pdf_doc = None try: @@ -527,7 +561,50 @@ def load_images_from_pdf_core( return images_list +def _model_input_bbox(item: dict[str, Any]) -> IntBBox: + """把模型输入项的 bbox 向外取整为裁图整数坐标。""" + bbox = item["bbox"] + assert bbox is not None + xmin, ymin, xmax, ymax = [float(value) for value in bbox] + return math.floor(xmin), math.floor(ymin), math.ceil(xmax), math.ceil(ymax) + + +def crop_img( + input_res: dict[str, Any], + input_img: Image.Image | np.ndarray, + crop_paste_x: int = 0, + crop_paste_y: int = 0, +) -> tuple[Image.Image | np.ndarray, list[int]]: + """按模型 bbox 裁图并在四周补白，返回坐标回投所需参数。""" + crop_xmin, crop_ymin, crop_xmax, crop_ymax = _model_input_bbox(input_res) + crop_new_width = crop_xmax - crop_xmin + crop_paste_x * 2 + crop_new_height = crop_ymax - crop_ymin + crop_paste_y * 2 + if isinstance(input_img, np.ndarray): + return_image: Image.Image | np.ndarray = np.ones((crop_new_height, crop_new_width, 3), dtype=np.uint8) * 255 + cropped_img = input_img[crop_ymin:crop_ymax, crop_xmin:crop_xmax] + return_image[ + crop_paste_y : crop_paste_y + (crop_ymax - crop_ymin), + crop_paste_x : crop_paste_x + (crop_xmax - crop_xmin), + ] = cropped_img + else: + return_image = Image.new("RGB", (crop_new_width, crop_new_height), "white") + cropped_img = input_img.crop((crop_xmin, crop_ymin, crop_xmax, crop_ymax)) + return_image.paste(cropped_img, (crop_paste_x, crop_paste_y)) + useful_list = [ + crop_paste_x, + crop_paste_y, + crop_xmin, + crop_ymin, + crop_xmax, + crop_ymax, + crop_new_width, + crop_new_height, + ] + return return_image, useful_list + + def get_crop_img(bbox: BBox, pil_img: Image.Image, scale: float = 2.0) -> Image.Image: + """按缩放 bbox 裁剪 Pillow 图像。""" scale_bbox = normalize_to_int_bbox([float(v) * scale for v in bbox]) if scale_bbox is None: return pil_img.crop((0, 0, 0, 0)) @@ -535,6 +612,7 @@ def get_crop_img(bbox: BBox, pil_img: Image.Image, scale: float = 2.0) -> Image. def get_crop_np_img(bbox: BBox, input_img: Image.Image | np.ndarray, scale: float = 2.0) -> np.ndarray: + """按缩放 bbox 裁剪 Pillow 或 NumPy 图像并返回数组。""" if isinstance(input_img, Image.Image): np_img = np.asarray(input_img) elif isinstance(input_img, np.ndarray): @@ -553,28 +631,13 @@ def get_crop_np_img(bbox: BBox, input_img: Image.Image | np.ndarray, scale: floa return np_img[scale_bbox[1] : scale_bbox[3], scale_bbox[0] : scale_bbox[2]] -def images_bytes_to_pdf_bytes(image_bytes: bytes) -> bytes: - # 内存缓冲区 - pdf_buffer = BytesIO() - - # 载入并转换所有图像为 RGB 模式 - image = Image.open(BytesIO(image_bytes)) - # 根据 EXIF 信息自动转正（处理手机拍摄的带 Orientation 标记的图片） - image = ImageOps.exif_transpose(image) or image - # 只在必要时转换 - if image.mode != "RGB": - image = image.convert("RGB") - - # 第一张图保存为 PDF，其余追加 - image.save( - pdf_buffer, - format="PDF", - resolution=DEFAULT_PDF_IMAGE_DPI, - quality=95, - subsampling=0, - ) - - # 获取 PDF bytes 并重置指针（可选） - pdf_bytes = pdf_buffer.getvalue() - pdf_buffer.close() - return pdf_bytes +__all__ = [ + "ImageType", + "crop_img", + "get_crop_img", + "get_crop_np_img", + "load_images_from_pdf_bytes_range", + "load_images_from_pdf_core", + "pdf_page_to_image", + "shutdown_pdf_render_executor", +] diff --git a/mineru/backend/analysis/pdf/layout.py b/mineru/backend/analysis/pdf/layout.py index a1f3b182..b6714f60 100644 --- a/mineru/backend/analysis/pdf/layout.py +++ b/mineru/backend/analysis/pdf/layout.py @@ -10,9 +10,9 @@ from typing import Any import numpy as np from PIL import Image -from mineru.types import BBox, BlockType -from mineru.utils.bbox_utils import normalize_to_int_bbox -from mineru.utils.pdf_image_tools import get_crop_np_img +from ....types import BBox, BlockType +from ....utils.geometry import normalize_to_int_bbox +from .images import get_crop_np_img from .geometry import _normalize_layout_bbox_to_unit, _normalize_page_size from .constants import ( @@ -24,7 +24,7 @@ from .constants import ( def _load_vlm_runtime() -> dict[str, Any]: """按需加载 VLM runtime 组件，确保只有 high/extra_high 路径触发 VLM 依赖。""" - from mineru.model.vlm.runtime import ( + from ....model.vlm.runtime import ( ModelSingleton, _get_model_async, _maybe_enable_serial_execution, diff --git a/mineru/backend/analysis/pdf/normalization.py b/mineru/backend/analysis/pdf/normalization.py index 0b3d7c7f..c0c0c395 100644 --- a/mineru/backend/analysis/pdf/normalization.py +++ b/mineru/backend/analysis/pdf/normalization.py @@ -8,9 +8,9 @@ import re from collections.abc import Callable from typing import Any -from mineru.types import BBox, BlockType, RAW_PHONETIC -from mineru.utils.bbox_utils import calculate_overlap_area_2_minbox_area_ratio -from mineru.utils.text_utils import full_to_half_exclude_marks +from ....types import BBox, BlockType, RAW_PHONETIC +from ....utils.geometry import calculate_overlap_area_2_minbox_area_ratio +from ....utils.text import full_to_half_exclude_marks from .constants import ( LAYOUT_TITLE_SPLIT_OVERLAP_THRESHOLD, diff --git a/mineru/backend/analysis/pdf/ocr.py b/mineru/backend/analysis/pdf/ocr.py index 792f4114..270d0e27 100644 --- a/mineru/backend/analysis/pdf/ocr.py +++ b/mineru/backend/analysis/pdf/ocr.py @@ -10,19 +10,13 @@ import cv2 import numpy as np from tqdm import tqdm -from mineru.backend.local_model_runtime import HybridLocalModelContext -from mineru.types import BlockType -from mineru.utils.bbox_utils import normalize_to_int_bbox -from mineru.utils.model_utils import crop_img -from mineru.utils.ocr_utils import ( - OcrConfidence, - get_adjusted_mfdetrec_res, - get_ocr_result_list, - mask_formula_regions_for_ocr_det, - merge_det_boxes, - sorted_boxes, - update_det_boxes, -) +from ....model.runtime.hybrid import HybridLocalModelContext +from ....types import BlockType +from ....utils.geometry import normalize_to_int_bbox +from .images import crop_img +from ....model.ocr.geometry import merge_det_boxes, sorted_boxes, update_det_boxes +from ....model.ocr.image import mask_formula_regions_for_ocr_det +from ....model.ocr.results import OcrConfidence, get_adjusted_mfdetrec_res, get_ocr_result_list from .constants import ( BATCH_RATIO, diff --git a/mineru/backend/analysis/pdf/pipeline.py b/mineru/backend/analysis/pdf/pipeline.py index 8ab0e186..4709b23a 100644 --- a/mineru/backend/analysis/pdf/pipeline.py +++ b/mineru/backend/analysis/pdf/pipeline.py @@ -3,22 +3,19 @@ from __future__ import annotations -import os import time from typing import Any, cast -from mineru.backend.analysis.contracts import AnalysisResult, AnalyzeEffort, ParseMode, ResolvedParseMode -from mineru.backend.local_model_runtime import HybridLocalModelContextSingleton -from mineru.utils.engine_utils import get_vlm_engine -from mineru.utils.model_utils import clean_memory -from mineru.utils.pdf_document import PDFDocument +from ..contracts import AnalysisResult, AnalyzeEffort, ParseMode, ResolvedParseMode +from ....model.runtime.hybrid import HybridLocalModelContextSingleton +from ....model.runtime.memory import clean_memory +from ....model.vlm.selector import get_vlm_engine +from ....model.flash.pdf.document import PDFDocument from .layout import _load_vlm_runtime from .normalization import _normalize_pdf_model_list from .window import process_pdf_windows -os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1" - _SUPPORTED_PDF_EFFORTS = {"flash", "medium", "high", "xhigh"} @@ -29,6 +26,10 @@ def analyze_pdf( image_analysis: bool = True, ) -> AnalysisResult: """生产 PDF model-list，并返回最终路由元数据和精确推理耗时。""" + # 只在真实 PDF 分析开始时配置 MPS 回退，避免 import backend 修改进程环境。 + import os + + os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1") if effort not in _SUPPORTED_PDF_EFFORTS: raise ValueError(f"Unsupported analyze effort: {effort}") diff --git a/mineru/backend/analysis/pdf/tables.py b/mineru/backend/analysis/pdf/tables.py index c8ffe020..0972642c 100644 --- a/mineru/backend/analysis/pdf/tables.py +++ b/mineru/backend/analysis/pdf/tables.py @@ -13,22 +13,22 @@ import cv2 import numpy as np from loguru import logger -from ...local_model_runtime import HybridLocalModelContext, run_ocr_inference -from ....model.model_types import AtomicModelName +from ....model.runtime.hybrid import HybridLocalModelContext, run_ocr_inference +from ....model.runtime.contracts import AtomicModelName from ....types import RAW_ALGORITHM, RAW_FORMULA_NUMBER, RAW_PHONETIC, BBox, BlockType -from ....utils.bbox_utils import ( +from ....utils.geometry import ( calculate_overlap_area_in_bbox1_area_ratio, normalize_to_int_bbox, ) -from ....utils.native_pdf_table import ( +from ....model.flash.pdf.table_recovery import ( NativeTableInput, coerce_native_table_rectangles, coerce_native_table_rules, recover_native_pdf_table, ) -from ....utils.ocr_utils import mask_formula_regions_for_ocr_det -from ....utils.pdf_document import PDFPage, get_lines_from_chars -from ....utils.spatial_text import project_ocr_table_text +from ....model.ocr.image import mask_formula_regions_for_ocr_det +from ....model.flash.pdf.document import PDFPage, get_lines_from_chars +from ....model.flash.pdf.spatial_text import project_ocr_table_text from .constants import ( BATCH_RATIO, diff --git a/mineru/backend/analysis/pdf/text/__init__.py b/mineru/backend/analysis/pdf/text/__init__.py index 20d5832e..86af3308 100644 --- a/mineru/backend/analysis/pdf/text/__init__.py +++ b/mineru/backend/analysis/pdf/text/__init__.py @@ -1,2 +1,4 @@ # Copyright (c) Opendatalab. All rights reserved. """PDF 原生文本、OCR Span 和内容重建。""" + +__all__: list[str] = [] diff --git a/mineru/backend/analysis/pdf/text/content.py b/mineru/backend/analysis/pdf/text/content.py index 0c5b2b76..78e9897a 100644 --- a/mineru/backend/analysis/pdf/text/content.py +++ b/mineru/backend/analysis/pdf/text/content.py @@ -9,18 +9,19 @@ import numpy as np from PIL import Image from pdftext.schema import Char -from mineru.backend.local_model_runtime import HybridLocalModelContext, run_ocr_inference -from mineru.types import BBox, BlockType, ContentType -from mineru.utils.language import detect_lang -from mineru.utils.ocr_utils import OcrConfidence, rotate_vertical_crop_if_needed -from mineru.utils.pdf_document import PDFPage, get_lines_from_chars -from mineru.utils.pdf_text_styles import ( +from .....model.runtime.hybrid import HybridLocalModelContext, run_ocr_inference +from .....types import BBox, BlockType, ContentType +from .....utils.language import detect_lang +from .....model.ocr.image import rotate_vertical_crop_if_needed +from .....model.ocr.results import OcrConfidence +from .....model.flash.pdf.document import PDFPage, get_lines_from_chars +from .....model.flash.pdf.text_styles import ( PDFTextLinkLine, PDFTextStyleLine, apply_pdf_text_links, apply_pdf_text_styles, ) -from mineru.utils.text_utils import merge_text_line_contents +from .....utils.text import merge_text_line_contents from ..constants import ( CODE_CONTENT_BLOCK_TYPES, diff --git a/mineru/backend/analysis/pdf/text/lines.py b/mineru/backend/analysis/pdf/text/lines.py index 8170ef55..e3eea1c7 100644 --- a/mineru/backend/analysis/pdf/text/lines.py +++ b/mineru/backend/analysis/pdf/text/lines.py @@ -4,7 +4,7 @@ from __future__ import annotations from .....types import ContentType -from .....utils.ocr_utils import _is_overlaps_x_exceeds_threshold, _is_overlaps_y_exceeds_threshold +from .....model.ocr.geometry import _is_overlaps_x_exceeds_threshold, _is_overlaps_y_exceeds_threshold from .models import _AnalyzeLine, _AnalyzeSpan VERTICAL_SPAN_HEIGHT_TO_WIDTH_RATIO_THRESHOLD = 2 diff --git a/mineru/backend/analysis/pdf/text/native.py b/mineru/backend/analysis/pdf/text/native.py index 7f398ddb..8aae32c5 100644 --- a/mineru/backend/analysis/pdf/text/native.py +++ b/mineru/backend/analysis/pdf/text/native.py @@ -17,10 +17,10 @@ from loguru import logger from pdftext.schema import Char from .....types import BBox, BlockType, ContentType -from .....utils.image_utils import calculate_contrast -from .....utils.pdf_document import PDFPage, get_lines_from_chars -from .....utils.pdf_image_tools import get_crop_img -from mineru.utils.bbox_utils import calculate_overlap_area_in_bbox1_area_ratio +from .....utils.image import calculate_contrast +from .....model.flash.pdf.document import PDFPage, get_lines_from_chars +from ..images import get_crop_img +from .....utils.geometry import calculate_overlap_area_in_bbox1_area_ratio from .models import _AnalyzeSpan MAX_NATIVE_TEXT_CHARS_PER_PAGE = 65535 diff --git a/mineru/backend/analysis/pdf/text/styles.py b/mineru/backend/analysis/pdf/text/styles.py index 968a1a81..1f66d4fe 100644 --- a/mineru/backend/analysis/pdf/text/styles.py +++ b/mineru/backend/analysis/pdf/text/styles.py @@ -7,8 +7,8 @@ from typing import Any, Sequence from pdftext.schema import Char -from mineru.utils.pdf_document import PDFPage, get_lines_from_chars -from mineru.utils.pdf_text_styles import ( +from .....model.flash.pdf.document import PDFPage, get_lines_from_chars +from .....model.flash.pdf.text_styles import ( PDFTextLinkLine, PDFTextStyleLine, detect_pdf_text_link_lines, @@ -30,7 +30,7 @@ def build_pdf_native_visual_lines_and_styles( """一次读取当前页字符，构造视觉 run、文本样式和超链接证据。""" # 延迟导入避免 Hybrid 模块初始化时提前加载完整 Flash PDF 流水线。 - from mineru.model.flash.native_pdf.native_text import _build_native_line_items + from .....model.flash.pdf.native_text import _build_native_line_items chars = page_chars if page_chars is not None else pdf_page.get_chars() line_items = _build_native_line_items( diff --git a/mineru/backend/analysis/pdf/visuals.py b/mineru/backend/analysis/pdf/visuals.py index 97ae70d9..84699b34 100644 --- a/mineru/backend/analysis/pdf/visuals.py +++ b/mineru/backend/analysis/pdf/visuals.py @@ -9,8 +9,8 @@ from typing import Any import numpy as np from loguru import logger -from mineru.types import BBox, BlockType -from mineru.utils.bbox_utils import ( +from ....types import BBox, BlockType +from ....utils.geometry import ( calculate_overlap_area_2_minbox_area_ratio, calculate_overlap_area_in_bbox1_area_ratio, ) diff --git a/mineru/backend/analysis/pdf/window.py b/mineru/backend/analysis/pdf/window.py index 013e44c0..c39763a3 100644 --- a/mineru/backend/analysis/pdf/window.py +++ b/mineru/backend/analysis/pdf/window.py @@ -9,10 +9,9 @@ from typing import Any, Literal import numpy as np from loguru import logger -from mineru.backend.local_model_runtime import HybridLocalModelContext -from mineru.utils.config_reader import get_processing_window_size -from mineru.utils.pdf_document import PDFDocument, PDFPage -from mineru.utils.pdf_image_tools import load_images_from_pdf_bytes_range +from ....model.runtime.hybrid import HybridLocalModelContext +from ....model.flash.pdf.document import PDFDocument, PDFPage +from .images import load_images_from_pdf_bytes_range from ..contracts import AnalyzeEffort from .constants import ( @@ -61,6 +60,20 @@ from .text.content import ( ) +def _configured_window_size(default: int = 64) -> int: + """读取 PDF 分析窗口大小，非法配置回退到正整数默认值。""" + import os + + raw_value = os.getenv("MINERU_PROCESSING_WINDOW_SIZE") + if raw_value is None: + return default + try: + return max(1, int(raw_value)) + except ValueError: + logger.warning(f"Invalid MINERU_PROCESSING_WINDOW_SIZE value: {raw_value}, use default {default}") + return default + + @dataclass(frozen=True) class _ProcessingWindow: """记录单个 Hybrid 处理窗口的页码范围。""" @@ -278,11 +291,11 @@ def process_pdf_windows( model_list: list[list[dict[str, Any]]] = [] if flash_txt_mode: # Flash 先对整份 PDF 生成完整 model_list，不依赖页面渲染和处理窗口。 - from mineru.model.flash import PdfModel + from ....model.flash import PdfModel model_list = PdfModel().predict(document) - configured_window_size = get_processing_window_size(default=64) + configured_window_size = _configured_window_size(default=64) windows = _build_processing_windows(page_count, configured_window_size) _log_processing_window_plan(page_count, configured_window_size, len(windows)) @@ -362,9 +375,7 @@ def process_pdf_windows( if effort == "medium": window_model_list = vl_style_layout_blocks elif effort == "high": - high_vlm_blocks, accepted_native_tables = _split_native_high_table_blocks( - vl_style_layout_blocks - ) + high_vlm_blocks, accepted_native_tables = _split_native_high_table_blocks(vl_style_layout_blocks) high_vlm_results = vlm_predictor.batch_extract_with_layout( images=images_pil_list, blocks_list=high_vlm_blocks, diff --git a/mineru/backend/analyze.py b/mineru/backend/analyze.py index 787452aa..25aa34ea 100644 --- a/mineru/backend/analyze.py +++ b/mineru/backend/analyze.py @@ -8,13 +8,10 @@ from typing import cast from loguru import logger -from mineru.backend.analysis.contracts import AnalyzeEffort, OfficeSuffix, ParseMode -from mineru.backend.analysis.office import analyze_office -from mineru.backend.analysis.pdf.pipeline import analyze_pdf -from mineru.backend.postprocess.document import model_json_to_middle_json -from mineru.config import config -from mineru.types import FILE_SUFFIXES, FileSuffix, MiddleJson, ModelJson -from mineru.version import __version__ as mineru_version +from .analysis.contracts import AnalyzeEffort, OfficeSuffix, ParseMode +from ..config import config +from ..types import FILE_SUFFIXES, FileSuffix, MiddleJson, ModelJson +from ..version import __version__ as mineru_version _SUPPORTED_ANALYZE_EFFORTS = {"flash", "medium", "high", "xhigh"} @@ -43,6 +40,8 @@ def doc_analyze( raise ValueError(f"Unsupported analyze effort: {effort}") if file_suffix == "pdf": + from .analysis.pdf.pipeline import analyze_pdf + result = analyze_pdf( file_bytes, effort=effort, @@ -50,6 +49,8 @@ def doc_analyze( image_analysis=image_analysis, ) else: + from .analysis.office import analyze_office + result = analyze_office(file_bytes, cast(OfficeSuffix, file_suffix)) _log_infer_performance(file_suffix, len(result.model_list), result.elapsed) @@ -61,6 +62,8 @@ def doc_analyze( parse_mode=result.parse_mode, mineru_version=mineru_version, ) + from .postprocess.document import model_json_to_middle_json + middle_json = model_json_to_middle_json( model_json, llm_aided_config=config.llm_aided, diff --git a/mineru/backend/postprocess/__init__.py b/mineru/backend/postprocess/__init__.py index 5b879f46..6562ef3d 100644 --- a/mineru/backend/postprocess/__init__.py +++ b/mineru/backend/postprocess/__init__.py @@ -1,2 +1,4 @@ # Copyright (c) Opendatalab. All rights reserved. """将严格 ModelJson 转换并增强为 MiddleJson 的后处理层。""" + +__all__: list[str] = [] diff --git a/mineru/backend/postprocess/document.py b/mineru/backend/postprocess/document.py index ba93ded1..be95d60f 100644 --- a/mineru/backend/postprocess/document.py +++ b/mineru/backend/postprocess/document.py @@ -3,8 +3,8 @@ from __future__ import annotations -from mineru.config import LLMAidedConfig -from mineru.types import MiddleJson, ModelJson +from ...config import LLMAidedConfig +from ...types import MiddleJson, ModelJson from .llm_aided import apply_llm_aided_postprocess from .pages import model_json_to_pages diff --git a/mineru/render/_internal/common/inline.py b/mineru/backend/postprocess/inline.py similarity index 98% rename from mineru/render/_internal/common/inline.py rename to mineru/backend/postprocess/inline.py index fb5615f4..82fad8c7 100644 --- a/mineru/render/_internal/common/inline.py +++ b/mineru/backend/postprocess/inline.py @@ -8,8 +8,8 @@ import re from dataclasses import dataclass from typing import TypeAlias, Union -from mineru.utils.language import detect_lang -from mineru.utils.text_utils import CJK_LANGS, resolve_text_line_boundary +from ...utils.language import detect_lang +from ...utils.text import CJK_LANGS, resolve_text_line_boundary _INLINE_START_RE = re.compile( r"<(?Peq|text|hyperlink|sup|sub|strong|b|em|i|s|u)(?P\s[^<>]*?)?>", diff --git a/mineru/backend/postprocess/lists.py b/mineru/backend/postprocess/lists.py index be2ffa21..b6e845e3 100644 --- a/mineru/backend/postprocess/lists.py +++ b/mineru/backend/postprocess/lists.py @@ -6,8 +6,8 @@ from __future__ import annotations import re from typing import Any -from mineru.types import BlockType -from mineru.utils.bbox_utils import calculate_overlap_area_in_bbox1_area_ratio +from ...types import BlockType +from ...utils.geometry import calculate_overlap_area_in_bbox1_area_ratio from .visual import _bbox_for_calculation diff --git a/mineru/backend/postprocess/llm_aided.py b/mineru/backend/postprocess/llm_aided.py index ca1bbefa..50cb5f00 100644 --- a/mineru/backend/postprocess/llm_aided.py +++ b/mineru/backend/postprocess/llm_aided.py @@ -7,8 +7,8 @@ import asyncio from loguru import logger -from mineru.config import LLMAidedConfig -from mineru.types import MiddleJson +from ...config import LLMAidedConfig +from ...types import MiddleJson from .llm_client import LLMAidedClient from .table_merge.llm_cell_merge import apply_llm_cross_page_cell_merge diff --git a/mineru/backend/postprocess/llm_client.py b/mineru/backend/postprocess/llm_client.py index 095e4ac9..d31b2377 100644 --- a/mineru/backend/postprocess/llm_client.py +++ b/mineru/backend/postprocess/llm_client.py @@ -12,7 +12,7 @@ import json_repair from loguru import logger from openai import AsyncOpenAI -from mineru.config import LLMAidedConfig +from ...config import LLMAidedConfig ValidatedResult = TypeVar("ValidatedResult") diff --git a/mineru/backend/postprocess/page_blocks.py b/mineru/backend/postprocess/page_blocks.py index b6cd1393..c4e52ac6 100644 --- a/mineru/backend/postprocess/page_blocks.py +++ b/mineru/backend/postprocess/page_blocks.py @@ -6,16 +6,16 @@ from __future__ import annotations import re from typing import Any -from mineru.types import ( +from ...types import ( RAW_ALGORITHM, RAW_CAPTION, RAW_FOOTNOTE, BlockType, VISUAL_MAIN_TYPES, ) -from mineru.utils.guess_suffix_or_lang import guess_language_by_text +from ...utils.language import guess_code_language -from mineru.utils.text_utils import clean_isolated_formula +from ...utils.text import clean_isolated_formula from .content import clean_content, code_content_clean from .lists import ( @@ -167,7 +167,7 @@ def _annotate_code_languages(code_blocks: list[BlockDict]) -> None: continue for sub_block in code_block["content"]: if sub_block.get("type") == BlockType.CODE_BODY: - code_block["guess_lang"] = guess_language_by_text(sub_block.get("content", "")) + code_block["guess_lang"] = guess_code_language(sub_block.get("content", "")) break diff --git a/mineru/backend/postprocess/pages.py b/mineru/backend/postprocess/pages.py index c104fcfb..67aef9db 100644 --- a/mineru/backend/postprocess/pages.py +++ b/mineru/backend/postprocess/pages.py @@ -8,11 +8,11 @@ from typing import Any from pydantic import TypeAdapter -from mineru.backend.postprocess.lists import fix_office_index_title_blocks, fix_office_paragraph_titles -from mineru.backend.postprocess.page_blocks import process_page_blocks -from mineru.backend.postprocess.paragraphs import merge_para_text_blocks -from mineru.backend.postprocess.table_merge import merge_table -from mineru.types import ModelJson, PageInfo +from .lists import fix_office_index_title_blocks, fix_office_paragraph_titles +from .page_blocks import process_page_blocks +from .paragraphs import merge_para_text_blocks +from .table_merge import merge_table +from ...types import ModelJson, PageInfo PAGE_INFO_LIST_ADAPTER = TypeAdapter(list[PageInfo]) diff --git a/mineru/backend/postprocess/table_merge/blocks.py b/mineru/backend/postprocess/table_merge/blocks.py index 49d9364b..211a60fe 100644 --- a/mineru/backend/postprocess/table_merge/blocks.py +++ b/mineru/backend/postprocess/table_merge/blocks.py @@ -8,7 +8,7 @@ from typing import Any from bs4 import BeautifulSoup -from mineru.types import BlockType +from ....types import BlockType from .html import _build_front_cache, _scan_rows from .models import ( diff --git a/mineru/backend/postprocess/table_merge/content.py b/mineru/backend/postprocess/table_merge/content.py index f4079ced..c7765070 100644 --- a/mineru/backend/postprocess/table_merge/content.py +++ b/mineru/backend/postprocess/table_merge/content.py @@ -7,7 +7,7 @@ from copy import deepcopy from bs4 import Tag -from mineru.types import BlockType +from ....types import BlockType from .blocks import ( _build_post_body_child_index, diff --git a/mineru/backend/postprocess/table_merge/document.py b/mineru/backend/postprocess/table_merge/document.py index 55676096..05c64945 100644 --- a/mineru/backend/postprocess/table_merge/document.py +++ b/mineru/backend/postprocess/table_merge/document.py @@ -5,7 +5,7 @@ from __future__ import annotations from typing import Any -from mineru.types import BlockType +from ....types import BlockType from .blocks import _get_or_create_table_state from .models import BlockDict, PageInfoDict, TableMergeState diff --git a/mineru/backend/postprocess/table_merge/html.py b/mineru/backend/postprocess/table_merge/html.py index 3fd44bc7..4db15e51 100644 --- a/mineru/backend/postprocess/table_merge/html.py +++ b/mineru/backend/postprocess/table_merge/html.py @@ -7,7 +7,7 @@ from typing import Any from bs4 import BeautifulSoup, Tag -from mineru.utils.text_utils import full_to_half +from ....utils.text import full_to_half from .models import ( MAX_HEADER_ROWS, diff --git a/mineru/backend/postprocess/table_merge/llm_cell_merge.py b/mineru/backend/postprocess/table_merge/llm_cell_merge.py index b5422730..3668f29d 100644 --- a/mineru/backend/postprocess/table_merge/llm_cell_merge.py +++ b/mineru/backend/postprocess/table_merge/llm_cell_merge.py @@ -8,7 +8,7 @@ import json from dataclasses import dataclass from typing import Literal, cast -from mineru.types import PageInfo, TableBlock, TableBodyBlock +from ....types import PageInfo, TableBlock, TableBodyBlock from .html import build_row_rendered_cell_segments, build_table_state_from_html from .structure import _expand_header_count_by_rowspan, detect_table_headers diff --git a/mineru/backend/postprocess/table_merge/rules.py b/mineru/backend/postprocess/table_merge/rules.py index f033755a..acb4365a 100644 --- a/mineru/backend/postprocess/table_merge/rules.py +++ b/mineru/backend/postprocess/table_merge/rules.py @@ -1,7 +1,7 @@ # Copyright (c) Opendatalab. All rights reserved. """跨页表格延续文本与 caption 的轻量判定规则。""" -from mineru.utils.text_utils import full_to_half +from ....utils.text import full_to_half CONTINUATION_END_MARKERS = [ "(续)", diff --git a/mineru/backend/postprocess/table_merge/structure.py b/mineru/backend/postprocess/table_merge/structure.py index 887e3745..4b48d83a 100644 --- a/mineru/backend/postprocess/table_merge/structure.py +++ b/mineru/backend/postprocess/table_merge/structure.py @@ -7,7 +7,7 @@ from typing import Any from bs4 import Tag -from mineru.types import BlockType +from ....types import BlockType from .blocks import ( _bbox_for_calculation, diff --git a/mineru/backend/postprocess/title_leveling.py b/mineru/backend/postprocess/title_leveling.py index 343d4891..23552b62 100644 --- a/mineru/backend/postprocess/title_leveling.py +++ b/mineru/backend/postprocess/title_leveling.py @@ -7,8 +7,8 @@ import asyncio import json from typing import Any -from mineru.types import DocTitleBlock, PageInfo, ParagraphTitleBlock -from mineru.render._internal.common.inline import inline_plain_text, parse_inline_content +from ...types import DocTitleBlock, PageInfo, ParagraphTitleBlock +from .inline import inline_plain_text, parse_inline_content from .llm_client import LLMAidedClient diff --git a/mineru/backend/postprocess/visual.py b/mineru/backend/postprocess/visual.py index 4c04d3a5..6c5e57ad 100644 --- a/mineru/backend/postprocess/visual.py +++ b/mineru/backend/postprocess/visual.py @@ -13,8 +13,8 @@ from ...types import ( VISUAL_RELATION_IGNORED_TYPES, VISUAL_TYPE_MAPPING, ) -from mineru.utils.bbox_utils import bbox_center_distance, bbox_distance -from mineru.types import RAW_CAPTION, RAW_FOOTNOTE +from ...utils.geometry import bbox_center_distance, bbox_distance +from ...types import RAW_CAPTION, RAW_FOOTNOTE from .table_merge.rules import is_table_continuation_text INLINE_CAPTION_FRAGMENT_TYPES = {BlockType.TEXT, RAW_FOOTNOTE} diff --git a/mineru/doclib/core/file_io.py b/mineru/doclib/core/file_io.py index 085fe82b..54bc86e1 100644 --- a/mineru/doclib/core/file_io.py +++ b/mineru/doclib/core/file_io.py @@ -9,7 +9,7 @@ from dataclasses import dataclass from pathlib import Path from ...filetypes import OFFICE_EXTENSIONS -from ...utils.pdf_document import PDFDocument +from ...model.flash.pdf.document import PDFDocument # Optional office doc support try: diff --git a/mineru/doclib/server.py b/mineru/doclib/server.py index 8cabbf56..021e610d 100644 --- a/mineru/doclib/server.py +++ b/mineru/doclib/server.py @@ -39,9 +39,9 @@ from ..types import ( _iter_child_blocks, select_highest_cached_tier, ) -from ..utils.model_registry import ModelRepo, model_repos_for_tier -from ..utils.models_download_utils import verify_model_repo -from ..utils.pdf_document import PDFDocument +from ..model.download import verify_model_repo +from ..model.registry import ModelRepo, model_repos_for_tier +from ..model.flash.pdf.document import PDFDocument from ..version import __version__ from .background.parse_server_health import get_health, get_managed_parse_server_tier from .base import AsyncDoclibInterface @@ -2078,7 +2078,7 @@ def _page_markdown_blocks( short_id: str = "", tier: Tier = "standard", ) -> list[tuple[int, str]]: - from mineru.config import config + from ..config import config result: list[tuple[int, str]] = [] image_renderer = _make_doclib_image_renderer( diff --git a/mineru/doclib/services/parse_svc.py b/mineru/doclib/services/parse_svc.py index c67677c1..c8832c06 100644 --- a/mineru/doclib/services/parse_svc.py +++ b/mineru/doclib/services/parse_svc.py @@ -1394,7 +1394,7 @@ class ParseService: if not pages: continue - from mineru.config import config + from ...config import config from ...render._internal.markdown.blocks import render_single_block delimiters = config.render.latex_delimiters diff --git a/mineru/kit/common.py b/mineru/kit/common.py index d8f0111e..9b8ad7ee 100644 --- a/mineru/kit/common.py +++ b/mineru/kit/common.py @@ -8,7 +8,7 @@ from typing import Literal, TypeAlias from ..filetypes import PARSEABLE_EXTENSIONS from ..parser.base import ParseResult -from ..render.writer import FileBasedDataWriter +from ..parser.writer import FileBasedDataWriter from ..types import Tier from ..utils.image_payload import validate_image_sidecar_path diff --git a/mineru/parser/base.py b/mineru/parser/base.py index 50ba091e..844ab21f 100644 --- a/mineru/parser/base.py +++ b/mineru/parser/base.py @@ -10,10 +10,10 @@ from typing import Any, Literal from ..render import render_markdown, render_structured_content from ..render.contracts import ImageRenderer, RenderMode -from ..render.writer import DataWriter +from .writer import DataWriter from ..types import FileSuffix, MiddleJson, PageInfo from ..utils.image_payload import ImagePayloadCache -from ..utils.pdf_document import PDFDocument +from ..model.flash.pdf.document import PDFDocument MIDDLE_JSON_SCHEMA_VERSION: str = "2.0" _LEGACY_SCHEMA_VERSION: str = "1.0" diff --git a/mineru/parser/mineru_parser.py b/mineru/parser/mineru_parser.py index cc1778c4..913cf54a 100644 --- a/mineru/parser/mineru_parser.py +++ b/mineru/parser/mineru_parser.py @@ -13,7 +13,7 @@ from ..backend.analyze import aio_doc_analyze, doc_analyze from ..errors import InvalidRequestError from ..filetypes import IMAGE_EXTENSIONS from ..types import FILE_SUFFIXES, FileSuffix, MiddleJson, ModelJson, PageInfo, Tier -from ..utils.backend_options import effort_for_tier +from .tier import effort_for_tier from .base import DocumentParser, ParseResult logger = logging.getLogger(__name__) @@ -111,8 +111,8 @@ class MinerUParser(DocumentParser): ) def _prepare_input(self, path: Path, page_range: str = "") -> _PreparedInput: - from ..utils.guess_suffix_or_lang import guess_suffix_by_path - from ..utils.pdf_document import PDFDocument + from .file_type import guess_suffix_by_path + from ..model.flash.pdf.document import PDFDocument file_name = path.stem file_bytes = path.read_bytes() @@ -163,8 +163,8 @@ class MinerUParser(DocumentParser): if suffix != "pdf": return file_bytes, None, None - from ..utils.pdf_document import PDFDocument - from ..utils.pdf_page_id import parse_page_range + from ..model.flash.pdf.document import PDFDocument + from .page_range import parse_page_range doc = PDFDocument(file_bytes) page_indices = parse_page_range(page_range, doc.page_count) @@ -178,7 +178,7 @@ class MinerUParser(DocumentParser): if page_indices == list(range(doc.page_count)): return file_bytes, None, None - from ..utils.pdfium_guard import safe_rewrite_pdf_bytes_with_pdfium_result + from ..model.flash.pdf.pdfium import safe_rewrite_pdf_bytes_with_pdfium_result rewrite_result = safe_rewrite_pdf_bytes_with_pdfium_result(file_bytes, page_indices=page_indices) if rewrite_result.used_original: diff --git a/mineru/render/_internal/__init__.py b/mineru/render/_internal/__init__.py index 070d526c..f512e1b1 100644 --- a/mineru/render/_internal/__init__.py +++ b/mineru/render/_internal/__init__.py @@ -1 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. """MinerU renderer 的非公共实现模块。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/common/__init__.py b/mineru/render/_internal/common/__init__.py index 4bcee000..5a0a176e 100644 --- a/mineru/render/_internal/common/__init__.py +++ b/mineru/render/_internal/common/__init__.py @@ -1 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. """各输出格式共用的渲染语义与规划实现。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/common/index.py b/mineru/render/_internal/common/index.py index 21142821..0538e612 100644 --- a/mineru/render/_internal/common/index.py +++ b/mineru/render/_internal/common/index.py @@ -5,7 +5,7 @@ from __future__ import annotations import re -from mineru.render._internal.common.inline import inline_plain_text, parse_inline_content +from ....backend.postprocess.inline import inline_plain_text, parse_inline_content _INDEX_ROMAN_RE = re.compile(r"[ivxlcdm]+", re.IGNORECASE) diff --git a/mineru/render/_internal/common/list_items.py b/mineru/render/_internal/common/list_items.py index 018bceb2..e6cb6427 100644 --- a/mineru/render/_internal/common/list_items.py +++ b/mineru/render/_internal/common/list_items.py @@ -7,8 +7,8 @@ from dataclasses import dataclass import re from typing import Literal, TypeAlias -from mineru.render._internal.common.inline import inline_plain_text, parse_inline_content -from mineru.types import BlockType, ListBlock +from ....backend.postprocess.inline import inline_plain_text, parse_inline_content +from ....types import BlockType, ListBlock ListItemKind: TypeAlias = Literal["unordered", "ordered", "explicit", "none"] OrderedListStyle: TypeAlias = Literal["decimal", "lower-alpha", "upper-alpha", "lower-roman", "upper-roman"] diff --git a/mineru/render/_internal/common/planner.py b/mineru/render/_internal/common/planner.py index ff7af5ce..9968b0ab 100644 --- a/mineru/render/_internal/common/planner.py +++ b/mineru/render/_internal/common/planner.py @@ -5,9 +5,9 @@ from __future__ import annotations from dataclasses import dataclass, field -from mineru.backend.postprocess.table_merge import merge_table_content -from mineru.render.contracts import RenderMode -from mineru.types import ( +from ....backend.postprocess.table_merge import merge_table_content +from ...contracts import RenderMode +from ....types import ( PAGE_AUXILIARY_BLOCK_TYPES, BlockType, ContinuableTextBlockBase, diff --git a/mineru/render/_internal/docx/__init__.py b/mineru/render/_internal/docx/__init__.py index 80c63481..c40522d2 100644 --- a/mineru/render/_internal/docx/__init__.py +++ b/mineru/render/_internal/docx/__init__.py @@ -1 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. """DOCX renderer 的内部实现。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/docx/assets.py b/mineru/render/_internal/docx/assets.py index e4b87d4e..8972ebd2 100644 --- a/mineru/render/_internal/docx/assets.py +++ b/mineru/render/_internal/docx/assets.py @@ -11,9 +11,9 @@ from urllib.parse import urlsplit from PIL import Image, UnidentifiedImageError -from mineru.render.contracts import AssetResolver -from mineru.types import ImagePayloadBlock -from mineru.utils.image_payload import parse_image_data_uri_strict, validate_image_sidecar_path +from ...contracts import AssetResolver +from ....types import ImagePayloadBlock +from ....utils.image_payload import parse_image_data_uri_strict, validate_image_sidecar_path _PIL_FORMAT_EXTENSIONS = { "BMP": "bmp", diff --git a/mineru/render/_internal/docx/inline.py b/mineru/render/_internal/docx/inline.py index f8f59e8e..4a9b116a 100644 --- a/mineru/render/_internal/docx/inline.py +++ b/mineru/render/_internal/docx/inline.py @@ -17,7 +17,7 @@ from docx.shared import Pt, RGBColor from lxml import etree from loguru import logger -from mineru.render._internal.common.inline import ( +from ....backend.postprocess.inline import ( InlineEquation, InlineLink, InlineNode, @@ -27,7 +27,7 @@ from mineru.render._internal.common.inline import ( join_inline_contents, parse_inline_content, ) -from mineru.render._internal.docx.math import DocxFormulaError, latex_to_omml +from .math import DocxFormulaError, latex_to_omml _BOOKMARK_SAFE_RE = re.compile(r"[^A-Za-z0-9_]+") _BOOKMARK_MAX_LENGTH = 40 diff --git a/mineru/render/_internal/docx/renderer.py b/mineru/render/_internal/docx/renderer.py new file mode 100644 index 00000000..941886e9 --- /dev/null +++ b/mineru/render/_internal/docx/renderer.py @@ -0,0 +1,852 @@ +# Copyright (c) Opendatalab. All rights reserved. +"""严格 MiddleJson 到可编辑 DOCX bytes 的公共渲染实现。""" + +from __future__ import annotations + +from collections.abc import Iterable +from io import BytesIO +import re + +from bs4 import BeautifulSoup, NavigableString, Tag +from docx import Document +from docx.document import Document as DocumentType +from docx.enum.table import WD_CELL_VERTICAL_ALIGNMENT +from docx.enum.text import WD_ALIGN_PARAGRAPH, WD_TAB_ALIGNMENT +from docx.oxml import OxmlElement +from docx.oxml.ns import qn +from docx.shared import Emu, Mm, Pt, Twips +from docx.table import _Cell +from docx.text.paragraph import Paragraph +from lxml import etree +from loguru import logger + +from ..common.index import strip_index_page_tail +from ....backend.postprocess.inline import ( + InlineEquation, + InlineLink, + InlineNode, + InlineStyled, + InlineText, + parse_inline_content, +) +from ..common.list_items import parse_list_item_marker +from ..common.planner import PlannedBlock, build_render_plan +from .assets import ( + DocxAssetError, + PreparedImage, + prepare_block_image, + prepare_html_image, +) +from .inline import ( + BookmarkRegistry, + InlineRenderContext, + append_inline_content, + append_inline_nodes, + append_internal_link, + append_joined_inline_contents, + sanitize_xml_text, +) +from .math import DocxFormulaError, latex_to_omml, split_formula_tag +from .styles import ( + AUXILIARY_STYLE, + BODY_STYLE, + CAPTION_STYLE, + CODE_STYLE, + FOOTNOTE_STYLE, + FORMULA_FALLBACK_STYLE, + SPATIAL_TABLE_STYLE, + configure_document, + usable_width_emu, + usable_width_twips, +) +from .table import DocxTableError, NestedTableWriter, materialize_docx_tables +from ...contracts import AssetResolver, RenderMode +from ...docx import DocxRenderError +from ....types import ( + PAGE_AUXILIARY_BLOCK_TYPES, + RAW_ALGORITHM, + BlockBase, + BlockType, + BBox, + ChartAnnotationBlock, + ChartBlock, + ChartBodyBlock, + CodeAnnotationBlock, + CodeBlock, + CodeBodyBlock, + DocTitleBlock, + EquationBlock, + ImageAnnotationBlock, + ImageBlock, + ImageBodyBlock, + ImagePayloadBlock, + IndexBlock, + ListBlock, + MiddleJson, + PageAuxTextBlock, + ParagraphTitleBlock, + RefTextBlock, + TableAnnotationBlock, + TableBlock, + TableBodyBlock, + TextBlock, + TitleBlockBase, +) + +_HTML_TABLE_RE = re.compile(r" None: + """初始化 renderer，并预注册全部标题 anchor。""" + self.middle_json = middle_json + self.mode = mode + self.asset_resolver = asset_resolver + self.document = Document() + configure_document(self.document) + self.bookmarks = BookmarkRegistry(_iter_document_anchors(middle_json)) + self.usable_width_emu = usable_width_emu(self.document) + self.usable_width_twips = usable_width_twips(self.document) + + def render(self) -> bytes: + """执行逐页 visitor，并把 python-docx document 序列化为 bytes。""" + planned_pages = build_render_plan(self.middle_json, self.mode) + for page_position, planned_blocks in enumerate(planned_pages): + if self.mode is RenderMode.FULL and page_position > 0: + self.document.add_page_break() + for planned in planned_blocks: + if planned.removed: + continue + if self.mode is RenderMode.DEFAULT and planned.block.type in PAGE_AUXILIARY_BLOCK_TYPES: + continue + self._render_planned_block(planned) + + output = BytesIO() + self.document.save(output) + return output.getvalue() + + def _render_planned_block(self, planned: PlannedBlock) -> None: + """按严格 PageBlock 具体类型分派 Word 写入逻辑。""" + block = planned.block + context = self._context(planned.page_idx, block) + if isinstance(block, (TextBlock, RefTextBlock)): + paragraph = self.document.add_paragraph(style=BODY_STYLE) + append_joined_inline_contents( + paragraph, + planned.text_contents or [block.content], + context=context, + ) + return + if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): + self._render_title(block, context) + return + if isinstance(block, PageAuxTextBlock): + paragraph = self.document.add_paragraph(style=AUXILIARY_STYLE) + append_inline_content(paragraph, block.content, context=context) + return + if isinstance(block, EquationBlock): + self._render_equation(block, context) + return + if isinstance(block, ListBlock): + self._render_list(block, context, depth=0) + return + if isinstance(block, IndexBlock): + self._render_index(block, context, depth=0) + return + if isinstance(block, ImageBlock): + self._render_image_block(block, context) + return + if isinstance(block, TableBlock): + self._render_table_block(block, context) + return + if isinstance(block, ChartBlock): + self._render_chart_block(block, context) + return + if isinstance(block, CodeBlock): + self._render_code_block(block, context) + return + raise TypeError(f"Unsupported PageBlock type: {type(block).__name__}") + + def _context(self, page_idx: int, block: BlockBase) -> InlineRenderContext: + """为当前 block 构造行内渲染上下文。""" + return InlineRenderContext( + bookmarks=self.bookmarks, + page_idx=page_idx, + block_index=block.index, + block_type=block.type, + ) + + def _render_title( + self, + block: DocTitleBlock | ParagraphTitleBlock, + context: InlineRenderContext, + ) -> None: + """用 Heading 1–6 写严格标题，并在段落内容外添加 bookmark。""" + level = min(max(block.level, 1), 9) + paragraph = self.document.add_paragraph(style=f"Heading {level}") + append_inline_content(paragraph, block.content, context=context) + self.bookmarks.attach(paragraph, block.anchor) + + def _render_equation(self, block: EquationBlock, context: InlineRenderContext) -> None: + """优先写可编辑 OMML，失败时按图片、可见 LaTeX 的顺序回退。""" + content = block.content.strip() + if content: + formula, tag = split_formula_tag(content) + try: + math_element = latex_to_omml(formula, display=True) + self._append_display_formula(math_element, tag, context=context) + return + except DocxFormulaError as exc: + logger.warning("DOCX display formula fallback: {} ({})", exc, context.location()) + + if block.image_base64 is not None or block.image_path is not None: + try: + self._append_block_image(block, context=context, alt_text="formula") + return + except DocxRenderError: + if not content: + raise + logger.warning("DOCX formula image fallback failed; preserving LaTeX ({})", context.location()) + + if content: + paragraph = self.document.add_paragraph(style=FORMULA_FALLBACK_STYLE) + paragraph.add_run(sanitize_xml_text(content, context=context)) + return + raise self._render_error("Equation does not contain usable LaTeX or image", context) + + def _append_display_formula( + self, + math_element: etree._Element, + tag: str | None, + *, + context: InlineRenderContext, + ) -> None: + """把块公式居中写入；带 tag 时使用中心与右对齐 tab。""" + if tag is None: + properties = OxmlElement("m:oMathParaPr") + justification = OxmlElement("m:jc") + justification.set(qn("m:val"), "center") + properties.append(justification) + math_element.insert(0, properties) + paragraph = self.document.add_paragraph(style=BODY_STYLE) + paragraph.paragraph_format.space_before = Pt(4) + paragraph.paragraph_format.space_after = Pt(5) + paragraph._p.append(math_element) + return + + equation = math_element.find(qn("m:oMath")) + if equation is None: + raise ValueError("Display OMML does not contain m:oMath") + math_element.remove(equation) + paragraph = self.document.add_paragraph(style=BODY_STYLE) + paragraph.paragraph_format.space_before = Pt(4) + paragraph.paragraph_format.space_after = Pt(5) + paragraph.paragraph_format.tab_stops.add_tab_stop( + Twips(self.usable_width_twips // 2), + WD_TAB_ALIGNMENT.CENTER, + ) + paragraph.paragraph_format.tab_stops.add_tab_stop( + Twips(self.usable_width_twips), + WD_TAB_ALIGNMENT.RIGHT, + ) + paragraph.add_run().add_tab() + paragraph._p.append(equation) + paragraph.add_run().add_tab() + rendered_tag = tag if tag.startswith("(") and tag.endswith(")") else f"({tag})" + paragraph.add_run(sanitize_xml_text(rendered_tag, context=context)) + + def _render_list( + self, + block: ListBlock, + context: InlineRenderContext, + *, + depth: int, + ) -> None: + """保留叶子原 marker，并仅用缩进表达递归列表层级。""" + for child in block.content: + if isinstance(child, ListBlock): + self._render_list(child, context, depth=depth + 1) + continue + paragraph = self.document.add_paragraph(style=BODY_STYLE) + item = parse_list_item_marker(child.content) + if item.marker is None: + paragraph.paragraph_format.left_indent = Mm(depth * 6) + else: + paragraph.paragraph_format.left_indent = Mm((depth + 1) * 6) + paragraph.paragraph_format.first_line_indent = Mm(-6) + append_inline_content(paragraph, child.content, context=context) + + def _render_index( + self, + block: IndexBlock, + context: InlineRenderContext, + *, + depth: int, + ) -> None: + """递归输出目录项，并把标题叶子链接到预注册 bookmark。""" + for child in block.content: + if isinstance(child, IndexBlock): + self._render_index(child, context, depth=depth + 1) + continue + content = strip_index_page_tail(child.content) + nodes = parse_inline_content(content) + if not nodes: + continue + paragraph = self.document.add_paragraph(style=BODY_STYLE) + paragraph.paragraph_format.left_indent = Mm((depth + 1) * 6) + paragraph.paragraph_format.first_line_indent = Mm(-4) + paragraph.add_run("• ") + anchor = child.anchor if isinstance(child, TitleBlockBase) else None + append_internal_link(paragraph, nodes, anchor=anchor, context=context) + + def _render_image_block(self, block: ImageBlock, context: InlineRenderContext) -> None: + """按原始子块顺序写图片主体、caption 与 footnote。""" + for child in block.content: + if isinstance(child, ImageBodyBlock): + alt_text = _plain_html_text(child.content) or block.sub_type or "image" + self._append_block_image(child, context=context, alt_text=alt_text) + elif isinstance(child, ImageAnnotationBlock): + self._render_annotation(child, context) + else: + raise TypeError(f"Unsupported image child: {type(child).__name__}") + + def _render_table_block(self, block: TableBlock, context: InlineRenderContext) -> None: + """按原始子块顺序写原生 HTML table 或预格式空间投影文本。""" + for child in block.content: + if isinstance(child, TableBodyBlock): + if _HTML_TABLE_RE.search(child.content): + try: + self._append_html_tables(child.content, context=context, depth=0) + except DocxRenderError as exc: + logger.warning("DOCX HTML table fallback: {}", exc) + if child.image_base64 is None and child.image_path is None: + raise self._render_error( + "HTML table cannot be materialized and has no image fallback", + context, + ) from exc + self._append_block_image(child, context=context, alt_text="table") + else: + if child.content and not child.content.isspace(): + paragraph = self.document.add_paragraph(style=SPATIAL_TABLE_STYLE) + paragraph.add_run(sanitize_xml_text(child.content, context=context)) + elif child.image_base64 is not None or child.image_path is not None: + self._append_block_image(child, context=context, alt_text="table") + else: + raise self._render_error( + "Spatial table does not contain text content or image", + context, + ) + elif isinstance(child, TableAnnotationBlock): + self._render_annotation(child, context) + else: + raise TypeError(f"Unsupported table child: {type(child).__name__}") + + def _render_chart_block(self, block: ChartBlock, context: InlineRenderContext) -> None: + """先写图表图片，再把 HTML 结构化数据追加为原生表格。""" + for child in block.content: + if isinstance(child, ChartBodyBlock): + has_image = child.image_base64 is not None or child.image_path is not None + if has_image: + self._append_block_image( + child, + context=context, + alt_text=_plain_html_text(child.content) or block.sub_type or "chart", + ) + if _HTML_TABLE_RE.search(child.content): + try: + self._append_html_tables(child.content, context=context, depth=0) + except DocxRenderError as exc: + if not has_image: + raise + logger.warning("DOCX chart data table omitted after image fallback: {}", exc) + elif child.content.strip() and not has_image: + paragraph = self.document.add_paragraph(style=BODY_STYLE) + append_inline_content(paragraph, child.content, context=context) + elif not has_image and not child.content.strip(): + raise self._render_error("Chart does not contain image or structured content", context) + elif isinstance(child, ChartAnnotationBlock): + self._render_annotation(child, context) + else: + raise TypeError(f"Unsupported chart child: {type(child).__name__}") + + def _render_code_block(self, block: CodeBlock, context: InlineRenderContext) -> None: + """按原始子块顺序写代码或算法主体及说明。""" + for child in block.content: + if isinstance(child, CodeBodyBlock): + paragraph = self.document.add_paragraph(style=CODE_STYLE) + if block.sub_type == BlockType.CODE: + paragraph.add_run(sanitize_xml_text(child.content, context=context)) + elif block.sub_type == RAW_ALGORITHM: + append_inline_content(paragraph, child.content, context=context) + else: + raise ValueError(f"Unsupported code subtype: {block.sub_type}") + elif isinstance(child, CodeAnnotationBlock): + self._render_annotation(child, context) + else: + raise TypeError(f"Unsupported code child: {type(child).__name__}") + + def _render_annotation( + self, + block: ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock, + context: InlineRenderContext, + ) -> None: + """根据说明类型选择 Caption 或 Footnote 样式并写入行内内容。""" + if block.type in _ANNOTATION_CAPTION_TYPES: + style = CAPTION_STYLE + elif block.type in _ANNOTATION_FOOTNOTE_TYPES: + style = FOOTNOTE_STYLE + else: + raise TypeError(f"Unsupported annotation type: {block.type}") + paragraph = self.document.add_paragraph(style=style) + append_inline_content(paragraph, block.content, context=context) + + def _append_block_image( + self, + block: ImagePayloadBlock, + *, + context: InlineRenderContext, + alt_text: str, + ) -> None: + """安全加载 block 图片，并按 bbox/自然尺寸限制到可用页面范围。""" + try: + prepared = prepare_block_image(block, self.asset_resolver) + except DocxAssetError as exc: + raise self._render_error(str(exc), context) from exc + self._append_prepared_image( + prepared, + bbox=block.bbox, + alt_text=alt_text, + max_width_emu=self.usable_width_emu, + context=context, + ) + + def _append_prepared_image( + self, + prepared: PreparedImage, + *, + bbox: BBox | None, + alt_text: str, + max_width_emu: int, + context: InlineRenderContext, + paragraph: Paragraph | None = None, + ) -> None: + """把准备好的图片按比例插入居中段落并写入 alt description。""" + target_paragraph = paragraph or self.document.add_paragraph(style=BODY_STYLE) + target_paragraph.alignment = WD_ALIGN_PARAGRAPH.CENTER + natural_width = round(prepared.width_px / 96 * 914400) + desired_width = natural_width + if bbox is not None: + desired_width = round(max_width_emu * max(0.0, min(1.0, bbox[2] - bbox[0]))) + minimum_width = min(int(Mm(5)), max_width_emu) + desired_width = max(minimum_width, min(desired_width, max_width_emu)) + + usable_height = int(Mm(257)) + rendered_height = round(desired_width * prepared.height_px / max(prepared.width_px, 1)) + if rendered_height > usable_height: + desired_width = round(usable_height * prepared.width_px / max(prepared.height_px, 1)) + + shape = target_paragraph.add_run().add_picture( + BytesIO(prepared.data), + width=Emu(max(1, desired_width)), + ) + description = sanitize_xml_text( + re.sub(r"\s+", " ", alt_text).strip()[:2048], + context=context, + ) + if description: + shape._inline.docPr.set("descr", description) + + def _append_html_tables( + self, + markup: str, + *, + context: InlineRenderContext, + depth: int, + container: DocumentType | _Cell | None = None, + ) -> None: + """物化 HTML 表格，并通过共享行内/素材逻辑填充 origin cell。""" + if depth >= 4: + raise self._render_error("Nested HTML table depth exceeds 4", context) + target_container = container or self.document + target_xml = target_container._element.body if isinstance(target_container, DocumentType) else target_container._tc + existing_child_count = len(target_xml) + insertion_index = existing_child_count - 1 if isinstance(target_container, DocumentType) else existing_child_count + existing_relationship_ids = set(self.document.part.rels) + + def fill_cell(cell: _Cell, source: Tag, write_nested: NestedTableWriter) -> None: + """填充一个 origin cell，并在遇到 nested table 时调用递归 writer。""" + self._fill_html_cell( + cell, + source, + write_nested=write_nested, + context=context, + ) + + try: + materialize_docx_tables( + target_container, + markup, + width_twips=self.usable_width_twips, + fill_cell=fill_cell, + ) + except DocxRenderError: + self._rollback_table_materialization( + target_xml, + existing_child_count=existing_child_count, + insertion_index=insertion_index, + existing_relationship_ids=existing_relationship_ids, + ) + raise + except DocxTableError as exc: + self._rollback_table_materialization( + target_xml, + existing_child_count=existing_child_count, + insertion_index=insertion_index, + existing_relationship_ids=existing_relationship_ids, + ) + raise self._render_error(str(exc), context) from exc + + def _rollback_table_materialization( + self, + target_xml: etree._Element, + *, + existing_child_count: int, + insertion_index: int, + existing_relationship_ids: set[str], + ) -> None: + """回滚本轮表格新增的 XML 根节点和 relationship，避免残留半成品。""" + while len(target_xml) > existing_child_count: + target_xml.remove(target_xml[insertion_index]) + for relationship_id in tuple(self.document.part.rels): + if relationship_id not in existing_relationship_ids: + self.document.part.drop_rel(relationship_id) + + def _fill_html_cell( + self, + cell: _Cell, + source: Tag, + *, + write_nested: NestedTableWriter, + context: InlineRenderContext, + ) -> None: + """按源顺序填充单元格文本、图片和任意包装层中的嵌套表格。""" + cell.text = "" + cell.vertical_alignment = WD_CELL_VERTICAL_ALIGNMENT.CENTER + _set_cell_margins(cell, top=100, start=120, bottom=100, end=120) + max_image_width_emu = _cell_content_width_emu(cell, horizontal_margin_twips=240) + paragraph = cell.paragraphs[0] + paragraph.paragraph_format.space_after = Pt(0) + paragraph.paragraph_format.line_spacing = 1.0 + + for child in source.children: + if not isinstance(child, (NavigableString, Tag)): + continue + if isinstance(child, Tag) and child.name.lower() in {"p", "div"} and _paragraph_has_content(paragraph): + paragraph = cell.add_paragraph() + paragraph.paragraph_format.space_after = Pt(0) + paragraph.paragraph_format.line_spacing = 1.0 + styles = ("bold",) if source.name == "th" else () + paragraph, _ = self._append_html_node( + paragraph, + child, + cell=cell, + write_nested=write_nested, + context=context, + inherited_styles=styles, + max_image_width_emu=max_image_width_emu, + ) + + def _append_html_node( + self, + paragraph: Paragraph, + node: NavigableString | Tag, + *, + cell: _Cell, + write_nested: NestedTableWriter, + context: InlineRenderContext, + inherited_styles: tuple[str, ...], + max_image_width_emu: int, + ) -> tuple[Paragraph, bool]: + """按 HTML 源顺序写入单元格行内文本、样式、公式和图片。""" + if isinstance(node, NavigableString): + text = str(node) + if not text: + return paragraph, False + nodes: list[InlineNode] = [InlineText(text)] + if inherited_styles: + nodes = [InlineStyled(nodes, inherited_styles)] + append_inline_nodes(paragraph, nodes, context=context) + return paragraph, bool(text.strip()) + + name = node.name.lower() + if name == "table": + write_nested(node) + trailing_paragraph = cell.paragraphs[-1] + trailing_paragraph.paragraph_format.space_after = Pt(0) + trailing_paragraph.paragraph_format.line_spacing = 1.0 + return trailing_paragraph, True + if name == "img": + self._append_html_image( + paragraph, + str(node.get("src", "")), + context=context, + alt_text=str(node.get("alt", "")), + max_width_emu=max_image_width_emu, + ) + return paragraph, True + if name == "br": + paragraph.add_run().add_break() + return paragraph, True + if name == "eq": + latex = node.get_text().strip() + nodes = [InlineEquation(latex)] if latex else [] + if inherited_styles and nodes: + nodes = [InlineStyled(nodes, inherited_styles)] + append_inline_nodes(paragraph, nodes, context=context) + return paragraph, bool(nodes) + if name == "a": + children = _html_inline_nodes(node) + if inherited_styles and children: + children = [InlineStyled(children, inherited_styles)] + append_inline_nodes( + paragraph, + [InlineLink(children, str(node.get("href", "")).strip())], + context=context, + ) + return paragraph, bool(children) + if name in {"ul", "ol"}: + start_value = node.get("start", "1") + try: + ordered_number = int(str(start_value)) + except ValueError: + ordered_number = 1 + rendered = False + for item in node.find_all("li", recursive=False): + if _paragraph_has_content(paragraph) or rendered: + paragraph.add_run().add_break() + marker = f"{ordered_number}. " if name == "ol" else "- " + paragraph.add_run(marker) + if name == "ol": + ordered_number += 1 + for child in item.children: + if isinstance(child, (NavigableString, Tag)): + paragraph, child_rendered = self._append_html_node( + paragraph, + child, + cell=cell, + write_nested=write_nested, + context=context, + inherited_styles=inherited_styles, + max_image_width_emu=max_image_width_emu, + ) + rendered = child_rendered or rendered + return paragraph, rendered + if name == "li": + rendered = False + for child in node.children: + if isinstance(child, (NavigableString, Tag)): + paragraph, child_rendered = self._append_html_node( + paragraph, + child, + cell=cell, + write_nested=write_nested, + context=context, + inherited_styles=inherited_styles, + max_image_width_emu=max_image_width_emu, + ) + rendered = child_rendered or rendered + return paragraph, rendered + if name == "text": + nodes = parse_inline_content(str(node)) + if inherited_styles and nodes: + nodes = [InlineStyled(nodes, inherited_styles)] + append_inline_nodes(paragraph, nodes, context=context) + return paragraph, bool(nodes) + + style = { + "strong": "bold", + "b": "bold", + "em": "italic", + "i": "italic", + "u": "underline", + "s": "strikethrough", + "sup": "superscript", + "sub": "subscript", + }.get(name) + styles = tuple(dict.fromkeys((*inherited_styles, style))) if style else inherited_styles + rendered = False + for child in node.children: + if isinstance(child, (NavigableString, Tag)): + paragraph, child_rendered = self._append_html_node( + paragraph, + child, + cell=cell, + write_nested=write_nested, + context=context, + inherited_styles=styles, + max_image_width_emu=max_image_width_emu, + ) + rendered = child_rendered or rendered + return paragraph, rendered + + def _append_html_image( + self, + paragraph: Paragraph, + source: str, + *, + context: InlineRenderContext, + alt_text: str, + max_width_emu: int, + ) -> None: + """安全加载表格单元格 img，并限制到紧凑的单元格宽度。""" + try: + prepared = prepare_html_image(source, self.asset_resolver) + except DocxAssetError as exc: + raise self._render_error(str(exc), context) from exc + self._append_prepared_image( + prepared, + bbox=None, + alt_text=alt_text or "table image", + max_width_emu=min(self.usable_width_emu, max_width_emu), + context=context, + paragraph=paragraph, + ) + + def _render_error(self, message: str, context: InlineRenderContext) -> DocxRenderError: + """用当前 context 构造带 page/block 定位的公共异常。""" + return DocxRenderError( + message, + page_idx=context.page_idx, + block_index=context.block_index, + block_type=context.block_type, + ) + + +def render_docx( + middle_json: MiddleJson, + *, + mode: RenderMode = RenderMode.DEFAULT, + asset_resolver: AssetResolver | None = None, +) -> bytes: + """把严格 MiddleJson 无副作用地渲染为完整 DOCX bytes。""" + if not isinstance(middle_json, MiddleJson): + raise TypeError("render_docx expects a MiddleJson instance") + if not isinstance(mode, RenderMode): + raise TypeError("mode must be a RenderMode value") + return _DocxRenderer( + middle_json, + mode=mode, + asset_resolver=asset_resolver, + ).render() + + +def _iter_document_anchors(middle_json: MiddleJson) -> Iterable[str]: + """按文档顺序遍历实际会写入 bookmark 的正文标题 anchor。""" + for page in middle_json.pages: + for block in page.blocks: + if isinstance(block, TitleBlockBase) and block.anchor: + yield block.anchor + + +def _plain_html_text(content: str) -> str: + """把图片或图表 body 内容收敛为适合 alt description 的纯文本。""" + if not content.strip(): + return "" + return BeautifulSoup(content, "html.parser").get_text(" ", strip=True) + + +def _html_inline_nodes(node: Tag) -> list[InlineNode]: + """把表格单元格中的安全 HTML 行内标签转换为共享 InlineNode。""" + nodes: list[InlineNode] = [] + for child in node.children: + if isinstance(child, NavigableString): + if str(child): + nodes.append(InlineText(str(child))) + continue + if not isinstance(child, Tag): + continue + name = child.name.lower() + children = _html_inline_nodes(child) + if name == "eq": + latex = child.get_text().strip() + if latex: + nodes.append(InlineEquation(latex)) + elif name in {"strong", "b"}: + nodes.append(InlineStyled(children, ("bold",))) + elif name in {"em", "i"}: + nodes.append(InlineStyled(children, ("italic",))) + elif name == "u": + nodes.append(InlineStyled(children, ("underline",))) + elif name == "s": + nodes.append(InlineStyled(children, ("strikethrough",))) + elif name == "sup": + nodes.append(InlineStyled(children, ("superscript",))) + elif name == "sub": + nodes.append(InlineStyled(children, ("subscript",))) + elif name == "a": + nodes.append(InlineLink(children, str(child.get("href", "")).strip())) + elif name == "br": + nodes.append(InlineText("\n")) + elif name == "text": + nodes.extend(parse_inline_content(str(child))) + elif name not in {"img", "table"}: + nodes.extend(children) + return nodes + + +def _set_cell_margins( + cell: _Cell, + *, + top: int, + start: int, + bottom: int, + end: int, +) -> None: + """用 DXA 为 Word 单元格写入四边内边距。""" + cell_properties = cell._tc.get_or_add_tcPr() + margins = cell_properties.find(qn("w:tcMar")) + if margins is None: + margins = OxmlElement("w:tcMar") + cell_properties.append(margins) + for side, value in (("top", top), ("start", start), ("bottom", bottom), ("end", end)): + element = margins.find(qn(f"w:{side}")) + if element is None: + element = OxmlElement(f"w:{side}") + margins.append(element) + element.set(qn("w:w"), str(value)) + element.set(qn("w:type"), "dxa") + + +def _cell_content_width_emu(cell: _Cell, *, horizontal_margin_twips: int) -> int: + """从当前物理或合并单元格 tcW 计算扣除左右内边距后的可用宽度。""" + cell_width = cell._tc.get_or_add_tcPr().get_or_add_tcW().w + width_twips = int(cell_width) if cell_width is not None else 1 + return int(Twips(max(1, width_twips - horizontal_margin_twips))) + + +def _paragraph_has_content(paragraph: Paragraph) -> bool: + """判断段落是否包含 pPr 之外的 run、公式、链接或 drawing。""" + return any(child.tag != qn("w:pPr") for child in paragraph._p) + + +__all__ = ["render_docx"] diff --git a/mineru/render/_internal/html/__init__.py b/mineru/render/_internal/html/__init__.py index 756f3dc3..7d71ccd6 100644 --- a/mineru/render/_internal/html/__init__.py +++ b/mineru/render/_internal/html/__init__.py @@ -1 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. """HTML renderer 的内部实现。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/html/inline.py b/mineru/render/_internal/html/inline.py index 499fe696..8c36c55a 100644 --- a/mineru/render/_internal/html/inline.py +++ b/mineru/render/_internal/html/inline.py @@ -9,7 +9,7 @@ import re from dataclasses import dataclass from urllib.parse import urlsplit -from mineru.render._internal.common.inline import ( +from ....backend.postprocess.inline import ( InlineEquation, InlineLink, InlineNode, @@ -19,25 +19,19 @@ from mineru.render._internal.common.inline import ( join_inline_contents, parse_inline_content, ) -from mineru.render._internal.html.sanitizer import sanitize_link_url +from .sanitizer import sanitize_link_url _AUTOLINK_TLD_PATTERN = r"(?:[A-Za-z]{2,63}|xn--[A-Za-z0-9-]{2,59})" -_AUTOLINK_DOMAIN_PATTERN = ( - rf"(?:[A-Za-z0-9](?:[A-Za-z0-9-]{{0,61}}[A-Za-z0-9])?\.){{1,10}}{_AUTOLINK_TLD_PATTERN}" -) +_AUTOLINK_DOMAIN_PATTERN = rf"(?:[A-Za-z0-9](?:[A-Za-z0-9-]{{0,61}}[A-Za-z0-9])?\.){{1,10}}{_AUTOLINK_TLD_PATTERN}" _AUTOLINK_EMAIL_EDGE_PATTERN = r"[A-Za-z0-9!#$%&'*+/=?^_{|}~-]" _AUTOLINK_EMAIL_PATTERN = ( rf"{_AUTOLINK_EMAIL_EDGE_PATTERN}" rf"(?:[A-Za-z0-9.!#$%&'*+/=?^_{{|}}~-]{{0,62}}{_AUTOLINK_EMAIL_EDGE_PATTERN})?" rf"@{_AUTOLINK_DOMAIN_PATTERN}" ) -_AUTOLINK_EXPLICIT_HOST_PATTERN = ( - rf"(?:{_AUTOLINK_DOMAIN_PATTERN}|localhost|(?:\d{{1,3}}\.){{3}}\d{{1,3}}|\[[0-9A-Fa-f:.]+\])" -) -_AUTOLINK_URL_TAIL_PATTERN = ( - r"(?::\d{1,5})?(?:[/?#](?:(?!https?://)[A-Za-z0-9._~:/?#\[\]@!$&()*+,;=%-])*)?" -) +_AUTOLINK_EXPLICIT_HOST_PATTERN = rf"(?:{_AUTOLINK_DOMAIN_PATTERN}|localhost|(?:\d{{1,3}}\.){{3}}\d{{1,3}}|\[[0-9A-Fa-f:.]+\])" +_AUTOLINK_URL_TAIL_PATTERN = r"(?::\d{1,5})?(?:[/?#](?:(?!https?://)[A-Za-z0-9._~:/?#\[\]@!$&()*+,;=%-])*)?" _AUTOLINK_RE = re.compile( rf"(?:(?P(?https?://{_AUTOLINK_EXPLICIT_HOST_PATTERN}{_AUTOLINK_URL_TAIL_PATTERN})" diff --git a/mineru/render/_internal/html/renderer.py b/mineru/render/_internal/html/renderer.py new file mode 100644 index 00000000..d6aeab17 --- /dev/null +++ b/mineru/render/_internal/html/renderer.py @@ -0,0 +1,941 @@ +# Copyright (c) Opendatalab. All rights reserved. +"""严格 MiddleJson 到 MinerU 风格 HTML 的公共渲染实现。""" + +from __future__ import annotations + +import html +import re +from functools import lru_cache +from importlib import resources +from urllib.parse import quote + +from bs4 import BeautifulSoup +from loguru import logger + +from ..common.index import strip_index_page_tail +from ....backend.postprocess.inline import inline_plain_text, parse_inline_content +from ..common.list_items import ListItem, parse_list_item_marker, reference_list_needs_bullets +from ..common.planner import PlannedBlock, build_render_plan +from .inline import ( + HtmlInlineResult, + render_inline_content_html, + render_inline_nodes_html, + render_joined_inline_contents_html, + render_math_html, +) +from .sanitizer import ( + is_supported_html_markup, + sanitize_html_fragment, + sanitize_image_source, +) +from .table import looks_like_gfm_table, render_gfm_table_html +from ...contracts import RenderMode +from ....types import ( + PAGE_AUXILIARY_BLOCK_TYPES, + RAW_ALGORITHM, + BlockType, + ChartAnnotationBlock, + ChartBlock, + ChartBodyBlock, + CodeAnnotationBlock, + CodeBlock, + CodeBodyBlock, + DocTitleBlock, + EquationBlock, + ImageAnnotationBlock, + ImageBlock, + ImageBodyBlock, + IndexBlock, + ListBlock, + MiddleJson, + PageAuxTextBlock, + ParagraphTitleBlock, + RefTextBlock, + TableAnnotationBlock, + TableBlock, + TableBodyBlock, + TextBlock, + TitleBlockBase, +) + +_STYLE_RESOURCE_NAME = "mineru.min.css" +_MATHJAX_URL = "https://cdn.jsdelivr.net/npm/mathjax@4.1.2/tex-chtml.js" +_MATHJAX_INTEGRITY = "sha384-zAhQQhdaMeHsMProNntGGg6nOUVcfuF9F22C3d1qJ9NZAVzCplXk1X85D2O5iufn" +_PRISM_CORE_URL = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/components/prism-core.min.js" +_PRISM_CORE_INTEGRITY = "sha384-zLRFO4dwowZvh8kzutOb5AWhH7f39HeJp+N7PtHF1SQtTBnifRx0AtmvTYs3F4YV" +_PRISM_AUTOLOADER_URL = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/plugins/autoloader/prism-autoloader.min.js" +_PRISM_AUTOLOADER_INTEGRITY = "sha384-Uq05+JLko69eOiPr39ta9bh7kld5PKZoU+fF7g0EXTAriEollhZ+DrN8Q/Oi8J2Q" +_PRISM_LANGUAGES_PATH = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/components/" +_MERMAID_URL = "https://cdn.jsdelivr.net/npm/mermaid@11.16.1/dist/mermaid.min.js" +_MERMAID_INTEGRITY = "sha384-aBQXj4hK6Jm05i7aQAsUV3bLdSUrHX1BGYfMB0166TtWt/RRaw+h0Eelme9OCOvy" +_MERMAID_MAX_TEXT_SIZE = 50_000 +_MERMAID_FLOWCHART_HEADER_RE = re.compile( + r"(?:graph|flowchart)\s+(?:TB|TD|BT|RL|LR)\b", + re.IGNORECASE, +) +_SAFE_PRISM_LANGUAGE_RE = re.compile(r"[a-z0-9][a-z0-9-]{0,31}") +_BLOCKED_PRISM_LANGUAGE_NAMES = {"constructor", "prototype"} +_PRISM_LANGUAGE_ALIASES = { + "c#": "csharp", + "c++": "cpp", + "html": "markup", + "js": "javascript", + "py": "python", + "sh": "bash", + "shell": "bash", + "none": "", + "text": "", + "txt": "", + "ts": "typescript", + "xml": "markup", + "yml": "yaml", +} + + +class _HtmlRenderer: + """维护一次 HTML 渲染所需的 anchor、公式与代码资源状态。""" + + def __init__( + self, + middle_json: MiddleJson, + *, + mode: RenderMode, + asset_base_url: str, + standalone: bool, + document_title: str | None, + ) -> None: + """保存严格输入和调用选项，并预收集真实正文标题 anchor。""" + self.middle_json = middle_json + self.mode = mode + self.asset_base_url = asset_base_url + self.standalone = standalone + self.document_title = document_title + self.anchor_targets = _collect_title_anchor_ids(middle_json) + self.emitted_anchors: set[str] = set() + self.has_math = False + self.has_prism = False + self.has_mermaid = False + + def render(self) -> str: + """生成 fragment article，并按需包装完整 HTML 文档。""" + planned_pages = build_render_plan(self.middle_json, self.mode) + if self.mode is RenderMode.FULL: + body = self._render_full_pages(planned_pages) + else: + body = self._render_default_pages(planned_pages) + article = f'
\n{body}\n
' + if not self.standalone: + return article + # 最终依赖以真实 DOM 为准，避免回退分支或普通文本中的 class 字样误触发。 + article_soup = BeautifulSoup(article, "html.parser") + self.has_math = article_soup.select_one(".mineru-math") is not None + self.has_prism = any( + any(str(class_name).startswith("language-") for class_name in code.get("class", [])) + for code in article_soup.find_all("code") + if code.get_text() + ) + self.has_mermaid = article_soup.select_one(".mineru-flowchart") is not None + return self._render_standalone(article) + + def _render_default_pages(self, planned_pages: list[list[PlannedBlock]]) -> str: + """把 DEFAULT 计划展平为无页面边界的连续阅读内容。""" + rendered: list[str] = [] + for page in planned_pages: + for planned in page: + block = self._render_planned_block(planned) + if block: + rendered.append(block) + return "\n".join(rendered) + + def _render_full_pages(self, planned_pages: list[list[PlannedBlock]]) -> str: + """为 FULL 计划保留所有页面 section 和相邻页面分隔。""" + sections: list[str] = [] + for page in planned_pages: + page_idx = page[0].page_idx if page else None + if page_idx is None: + page_position = len(sections) + page_idx = self.middle_json.pages[page_position].page_idx + blocks = [rendered for planned in page if (rendered := self._render_planned_block(planned))] + rendered_blocks = "\n".join(blocks) + sections.append(f'
\n{rendered_blocks}\n
') + return '\n\n'.join(sections) + + def _render_planned_block(self, planned: PlannedBlock) -> str: + """过滤计划块、分派类型 visitor，并添加稳定来源元数据。""" + if planned.removed: + return "" + block = planned.block + if self.mode is RenderMode.DEFAULT and block.type in PAGE_AUXILIARY_BLOCK_TYPES: + return "" + content = self._render_block_content(planned) + if not content or not content.strip(): + return "" + attrs = [ + 'class="mineru-block"', + f'data-page-idx="{planned.page_idx}"', + f'data-block-type="{html.escape(str(block.type), quote=True)}"', + ] + if block.index is not None: + attrs.append(f'data-block-index="{block.index}"') + return f"
\n{content}\n
" + + def _render_block_content(self, planned: PlannedBlock) -> str: + """按严格 PageBlock 具体类型返回对应语义 HTML。""" + block = planned.block + if isinstance(block, TextBlock): + rendered = render_joined_inline_contents_html(planned.text_contents or [block.content]) + self._observe_inline(rendered) + return f'

{rendered.html}

' if rendered.html else "" + if isinstance(block, RefTextBlock): + rendered = render_joined_inline_contents_html(planned.text_contents or [block.content]) + self._observe_inline(rendered) + return f'

{rendered.html}

' if rendered.html else "" + if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): + return self._render_title(block) + if isinstance(block, PageAuxTextBlock): + rendered = render_inline_content_html(block.content) + self._observe_inline(rendered) + class_name = html.escape(str(block.type).replace("_", "-"), quote=True) + return f'
{rendered.html}
' if rendered.html else "" + if isinstance(block, EquationBlock): + return self._render_equation(block) + if isinstance(block, ListBlock): + return self._render_list(block) + if isinstance(block, IndexBlock): + return self._render_index(block) + if isinstance(block, ImageBlock): + return self._render_image_block(block) + if isinstance(block, TableBlock): + return self._render_table_block(block) + if isinstance(block, ChartBlock): + return self._render_chart_block(block) + if isinstance(block, CodeBlock): + return self._render_code_block(block) + raise TypeError(f"Unsupported PageBlock type: {type(block).__name__}") + + def _render_title(self, block: DocTitleBlock | ParagraphTitleBlock) -> str: + """渲染标题，并保证正文 anchor 只在首次出现时生成 id。""" + rendered = render_inline_content_html(block.content) + self._observe_inline(rendered) + if not rendered.html: + return "" + level = min(max(block.level, 1), 6) + attrs = [f'class="mineru-heading mineru-heading--{level}"', f'data-heading-level="{block.level}"'] + anchor = _anchor_key(block.anchor) + if anchor: + if anchor not in self.emitted_anchors: + anchor_id = self.anchor_targets.get(anchor) + if anchor_id: + attrs.append(f'id="{html.escape(anchor_id, quote=True)}"') + self.emitted_anchors.add(anchor) + else: + logger.warning("Duplicate HTML title anchor omitted: {!r}", anchor) + return f"{rendered.html}" + + def _render_equation(self, block: EquationBlock) -> str: + """优先输出行间公式，空 LaTeX 时回退到安全图片来源。""" + if block.content.strip(): + rendered = render_math_html(block.content, display=True) + self._observe_inline(rendered) + return rendered.html + source = self._safe_block_image_source(block) + return self._render_image(source, alt="formula", class_name="mineru-equation-image") if source else "" + + def _render_list(self, block: ListBlock) -> str: + """按直属 marker 类型渲染一层列表，并递归挂接嵌套列表。""" + parsed_leaves = [ + parse_list_item_marker(child.content) + for child in block.content + if not isinstance(child, ListBlock) and child.content.strip() + ] + add_reference_bullets = reference_list_needs_bullets(block) + container_tag, list_type, class_name = _classify_list(parsed_leaves, add_reference_bullets) + items: list[dict[str, object]] = [] + expected_value: int | None = None + + for child in block.content: + if isinstance(child, ListBlock): + nested = self._render_list(child) + if not nested: + continue + if not items: + items.append({"content": "", "attrs": ' class="mineru-list-item--orphan"', "nested": []}) + nested_items = items[-1]["nested"] + assert isinstance(nested_items, list) + nested_items.append(nested) + continue + + parsed = parse_list_item_marker(child.content) + item_content, marker = _list_item_content( + parsed, + add_reference_bullets, + explicit_markers=class_name == "mineru-list--explicit", + ) + rendered = render_inline_content_html(item_content) + self._observe_inline(rendered) + if not rendered.html and not marker: + items.append({"content": "", "attrs": ' class="mineru-list-item--markerless"', "nested": []}) + continue + + attrs: list[str] = [] + if class_name == "mineru-list--explicit": + attrs.append('class="mineru-list-item--explicit"') + if container_tag == "ol" and parsed.kind == "ordered" and parsed.value is not None: + if expected_value is None: + expected_value = parsed.value + if parsed.value != expected_value: + attrs.append(f'value="{parsed.value}"') + expected_value = parsed.value + 1 + marker_html = ( + f'{html.escape(_replace_html_controls(marker), quote=False)}' + if marker + else "" + ) + if class_name == "mineru-list--explicit" and not marker_html: + marker_html = '' + items.append( + { + "content": f'{marker_html}{rendered.html}', + "attrs": f" {' '.join(attrs)}" if attrs else "", + "nested": [], + } + ) + + rendered_items: list[str] = [] + for item in items: + nested = item["nested"] + assert isinstance(nested, list) + item_content = str(item["content"]) + if not item_content and not nested: + continue + rendered_items.append(f"{item_content}{''.join(str(value) for value in nested)}") + if not rendered_items: + return "" + + container_attrs = [f'class="mineru-list {class_name}"'] + if list_type: + container_attrs.append(f'type="{list_type}"') + if container_tag == "ol" and parsed_leaves and parsed_leaves[0].value not in (None, 1): + container_attrs.append(f'start="{parsed_leaves[0].value}"') + return f"<{container_tag} {' '.join(container_attrs)}>{''.join(rendered_items)}" + + def _render_index(self, block: IndexBlock) -> str: + """递归渲染目录，并仅链接到真实正文标题 anchor。""" + items: list[dict[str, object]] = [] + for child in block.content: + if isinstance(child, IndexBlock): + nested = self._render_index_list(child) + if not nested: + continue + if not items: + items.append({"content": "", "orphan": True, "nested": []}) + nested_items = items[-1]["nested"] + assert isinstance(nested_items, list) + nested_items.append(nested) + continue + rendered = self._render_index_leaf(child) + items.append({"content": rendered, "orphan": False, "nested": []}) + inner = _serialize_index_items(items) + return f'' if inner else "" + + def _render_index_list(self, block: IndexBlock) -> str: + """渲染一个嵌套 IndexBlock 为可挂接到父项的 ul。""" + items: list[dict[str, object]] = [] + for child in block.content: + if isinstance(child, IndexBlock): + nested = self._render_index_list(child) + if nested: + if not items: + items.append({"content": "", "orphan": True, "nested": []}) + nested_items = items[-1]["nested"] + assert isinstance(nested_items, list) + nested_items.append(nested) + continue + rendered = self._render_index_leaf(child) + items.append({"content": rendered, "orphan": False, "nested": []}) + inner = _serialize_index_items(items) + return f"
    {inner}
" if inner else "" + + def _render_index_leaf(self, block: TextBlock | TitleBlockBase) -> str: + """渲染目录叶子，并在 anchor 命中正文标题时生成内部链接。""" + content = strip_index_page_tail(block.content) + rendered = render_inline_content_html(content) + self._observe_inline(rendered) + if not rendered.html: + return "" + anchor = _anchor_key(block.anchor) if isinstance(block, TitleBlockBase) else "" + if anchor and (anchor_id := self.anchor_targets.get(anchor)): + href = quote(anchor_id, safe="-._~") + return f'{rendered.html}' + return rendered.html + + def _render_image_block(self, block: ImageBlock) -> str: + """按原始子块顺序渲染图片主体及重复说明。""" + parts: list[str] = [] + for child in block.content: + if isinstance(child, ImageBodyBlock): + body = self._render_image_body(block, child) + elif isinstance(child, ImageAnnotationBlock): + body = self._render_annotation(child) + else: + raise TypeError(f"Unsupported image child: {type(child).__name__}") + if body: + parts.append(body) + return f'
{"".join(parts)}
' if parts else "" + + def _render_image_body(self, parent: ImageBlock, block: ImageBodyBlock) -> str: + """渲染图片，并把并存的识别内容放入原生 details。""" + if parent.sub_type == "flowchart": + mermaid_source = _extract_mermaid_flowchart_source(block.content) + if mermaid_source is not None: + return self._render_flowchart_body(block, mermaid_source) + source = self._safe_block_image_source(block) + content = self._render_embedded_content(block.content, linkify_text=parent.sub_type != "flowchart") + alt = _plain_content_text(block.content) or parent.sub_type or "image" + parts = [self._render_image(source, alt=alt, class_name="mineru-image")] if source else [] + if content.html: + if source: + parts.append(self._render_details(content.html, parent.sub_type or "image content")) + else: + parts.append(content.html) + return "".join(parts) + + def _render_flowchart_body(self, block: ImageBodyBlock, mermaid_source: str) -> str: + """输出 Mermaid canvas，并保留 raster 与源码两级失败回退。""" + source = self._safe_block_image_source(block) + escaped_source = html.escape(_replace_html_controls(mermaid_source), quote=False) + fallback = self._render_image(source, alt="flowchart", class_name="mineru-flowchart-fallback") if source else "" + fallback_class = " mineru-flowchart--has-raster" if fallback else "" + details_open = "" if fallback else " open" + return ( + f'
' + '' + f"{fallback}
" + f'
' + "flowchart source" + f'
{escaped_source}
' + "
" + ) + + def _render_table_block(self, block: TableBlock) -> str: + """按原始子块顺序渲染表格主体、标题与脚注。""" + parts: list[str] = [] + for child in block.content: + if isinstance(child, TableBodyBlock): + body = self._render_table_body(child) + elif isinstance(child, TableAnnotationBlock): + body = self._render_annotation(child) + else: + raise TypeError(f"Unsupported table child: {type(child).__name__}") + if body: + parts.append(body) + return f'
{"".join(parts)}
' if parts else "" + + def _render_table_body(self, block: TableBodyBlock) -> str: + """优先保留安全 HTML table，空间文本和整体图片依次回退。""" + normalized_content = _replace_html_controls(block.content) + content = normalized_content.strip() + if content: + source_soup = BeautifulSoup(content, "html.parser") + if source_soup.find("table") is not None: + rendered = self._render_embedded_content(content) + if rendered.html and _contains_usable_table(rendered.html): + return rendered.html + source = self._safe_block_image_source(block) + if source: + return self._render_image(source, alt="table", class_name="mineru-table-image") + return _render_raw_fallback(content) + return f'
{html.escape(normalized_content, quote=False)}
' + source = self._safe_block_image_source(block) + return self._render_image(source, alt="table", class_name="mineru-table-image") if source else "" + + def _render_chart_block(self, block: ChartBlock) -> str: + """按原始子块顺序渲染 chart 图片、结构内容及说明。""" + parts: list[str] = [] + for child in block.content: + if isinstance(child, ChartBodyBlock): + body = self._render_chart_body(block, child) + elif isinstance(child, ChartAnnotationBlock): + body = self._render_annotation(child) + else: + raise TypeError(f"Unsupported chart child: {type(child).__name__}") + if body: + parts.append(body) + return f'
{"".join(parts)}
' if parts else "" + + def _render_chart_body(self, parent: ChartBlock, block: ChartBodyBlock) -> str: + """渲染 chart 图片，并将并存结构内容放入 details。""" + source = self._safe_block_image_source(block) + content = self._render_chart_content(block.content) + parts = [self._render_image(source, alt=parent.sub_type or "chart", class_name="mineru-chart-image")] if source else [] + if content.html: + if source: + parts.append(self._render_details(content.html, parent.sub_type or "chart content")) + else: + parts.append(content.html) + return "".join(parts) + + def _render_chart_content(self, content: str) -> HtmlInlineResult: + """按 HTML、严格 GFM 表格、普通行内内容的顺序渲染 chart content。""" + normalized = content.strip() + if not normalized: + return HtmlInlineResult("") + if is_supported_html_markup(normalized): + return self._render_embedded_content(normalized) + if looks_like_gfm_table(normalized): + rendered = render_gfm_table_html(normalized) + if rendered is None: + return HtmlInlineResult(_render_raw_fallback(content)) + self._observe_inline(rendered) + return rendered + rendered = render_inline_content_html(content) + self._observe_inline(rendered) + return rendered + + def _render_code_block(self, block: CodeBlock) -> str: + """按原始子块顺序渲染代码或算法及其说明。""" + parts: list[str] = [] + for child in block.content: + if isinstance(child, CodeBodyBlock): + body = self._render_code_body(block, child) + elif isinstance(child, CodeAnnotationBlock): + body = self._render_annotation(child) + else: + raise TypeError(f"Unsupported code child: {type(child).__name__}") + if body: + parts.append(body) + return f'
{"".join(parts)}
' if parts else "" + + def _render_code_body(self, parent: CodeBlock, block: CodeBodyBlock) -> str: + """代码使用 Prism class，算法使用共享 inline AST 保留空白。""" + if parent.sub_type == BlockType.CODE: + escaped = html.escape(_replace_html_controls(block.content), quote=False) + if not block.content: + return '
' + language = _normalize_prism_language(parent.guess_lang) + class_attr = f' class="language-{language}"' if language else "" + pre_class = f"mineru-code language-{language}" if language else "mineru-code" + self.has_prism = bool(language) or self.has_prism + return f'
{escaped}
' + if parent.sub_type == RAW_ALGORITHM: + rendered = render_inline_nodes_html( + parse_inline_content(block.content), + linkify_text=False, + separate_adjacent_math=True, + preserve_newlines=True, + ) + self._observe_inline(rendered) + return f'
{rendered.html}
' if rendered.html else "" + raise ValueError(f"Unsupported code subtype: {parent.sub_type}") + + def _render_annotation( + self, + block: ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock, + ) -> str: + """把视觉说明渲染为保持原类型的独立段落。""" + rendered = render_inline_content_html(block.content) + self._observe_inline(rendered) + if not rendered.html: + return "" + is_caption = str(block.type).endswith("caption") + role_class = "mineru-caption" if is_caption else "mineru-footnote" + type_class = html.escape(str(block.type).replace("_", "-"), quote=True) + return f'

{rendered.html}

' + + def _render_embedded_content(self, content: str, *, linkify_text: bool = True) -> HtmlInlineResult: + """安全处理 body 富 HTML；普通内容走共享 inline AST。""" + normalized = content.strip() + if not normalized: + return HtmlInlineResult("") + if not is_supported_html_markup(normalized): + rendered = render_inline_nodes_html(parse_inline_content(content), linkify_text=linkify_text) + self._observe_inline(rendered) + return rendered + sanitized = sanitize_html_fragment(normalized, asset_base_url=self.asset_base_url) + rendered = self._replace_safe_equations(sanitized) + self._observe_inline(rendered) + return rendered + + def _replace_safe_equations(self, markup: str) -> HtmlInlineResult: + """把 sanitizer 保留的 eq 标签替换为可信 MathJax carrier。""" + if not markup: + return HtmlInlineResult("") + soup = BeautifulSoup(markup, "html.parser") + has_math = False + for equation in list(soup.find_all("eq")): + rendered = render_math_html(equation.get_text(), display=False) + if not rendered.html: + equation.decompose() + continue + replacement = BeautifulSoup(rendered.html, "html.parser") + equation.replace_with(*list(replacement.contents)) + has_math = True + return HtmlInlineResult(str(soup), has_math) + + def _safe_block_image_source(self, block: EquationBlock | ImageBodyBlock | TableBodyBlock | ChartBodyBlock) -> str | None: + """解析 block 图片来源，并应用 HTML renderer 的 URL 安全策略。""" + if block.image_path: + return sanitize_image_source(block.image_path, asset_base_url=self.asset_base_url) + if block.image_base64: + return sanitize_image_source(block.image_base64) + return None + + def _render_image(self, source: str, *, alt: str, class_name: str) -> str: + """构造属性已转义的安全 img 元素。""" + return ( + f'' + ) + + def _render_details(self, content: str, summary: str) -> str: + """构造浏览器原生折叠详情，不引入额外脚本。""" + return ( + '
' + f"{html.escape(_replace_html_controls(summary), quote=False)}{content}
" + ) + + def _observe_inline(self, rendered: HtmlInlineResult) -> None: + """累计当前文档是否需要加载 MathJax。""" + self.has_math = rendered.has_math or self.has_math + + def _render_standalone(self, article: str) -> str: + """用固定 CSS 和按需 CDN 资源包装完整 HTML 文档。""" + title = html.escape( + _replace_html_controls(_resolve_document_title(self.middle_json, self.document_title)), + quote=False, + ) + styles = _load_html_styles() + dependencies = [] + if self.has_math: + dependencies.append(_mathjax_head()) + if self.has_prism: + dependencies.append(_prism_head()) + if self.has_mermaid: + dependencies.append(_mermaid_head()) + dependency_html = "\n".join(dependencies) + return ( + "\n" + '\n' + "\n" + '\n' + '\n' + f"{title}\n" + f"\n" + f"{dependency_html}\n" + "\n" + '\n' + f"{article}\n" + "\n" + "" + ) + + +def render_html( + middle_json: MiddleJson, + *, + mode: RenderMode = RenderMode.DEFAULT, + asset_base_url: str = "", + standalone: bool = True, + document_title: str | None = None, +) -> str: + """把严格 MiddleJson 无副作用地渲染为 HTML 文档或单根 fragment。""" + if not isinstance(middle_json, MiddleJson): + raise TypeError("render_html expects a MiddleJson instance") + if not isinstance(mode, RenderMode): + raise TypeError("mode must be a RenderMode value") + if not isinstance(asset_base_url, str): + raise TypeError("asset_base_url must be a string") + if not isinstance(standalone, bool): + raise TypeError("standalone must be a bool") + if document_title is not None and not isinstance(document_title, str): + raise TypeError("document_title must be a string or None") + return _HtmlRenderer( + middle_json, + mode=mode, + asset_base_url=asset_base_url, + standalone=standalone, + document_title=document_title, + ).render() + + +def _classify_list(items: list[ListItem], add_reference_bullets: bool) -> tuple[str, str | None, str]: + """根据直属 marker 选择原生列表类型或显式 marker 模式。""" + if add_reference_bullets: + return "ul", None, "mineru-list--reference" + if items and all(item.kind == "unordered" for item in items): + return "ul", None, "mineru-list--unordered" + if items and all(item.kind == "ordered" for item in items): + styles = {item.ordered_style for item in items} + if len(styles) == 1: + style = next(iter(styles)) + list_type = { + "lower-alpha": "a", + "upper-alpha": "A", + "lower-roman": "i", + "upper-roman": "I", + }.get(style or "") + return "ol", list_type, "mineru-list--ordered" + if items and all(item.kind == "none" for item in items): + return "ul", None, "mineru-list--unmarked" + return "ul", None, "mineru-list--explicit" + + +def _list_item_content( + item: ListItem, + add_reference_bullets: bool, + *, + explicit_markers: bool, +) -> tuple[str, str | None]: + """决定一个列表项应剥离、保留还是显式显示源 marker。""" + if add_reference_bullets: + if item.kind == "unordered": + return item.body, None + original = f"{item.leading}{item.marker or ''}{item.separator}{item.body}" + return original, None + if explicit_markers: + return item.body, item.marker + if item.kind in {"unordered", "ordered"}: + return item.body, None + return item.body, item.marker + + +def _serialize_index_items(items: list[dict[str, object]]) -> str: + """把目录项及其已挂接的嵌套 ul 序列化为 li。""" + rendered: list[str] = [] + for item in items: + nested = item["nested"] + assert isinstance(nested, list) + content = str(item["content"]) + if not content and not nested: + continue + class_attr = ' class="mineru-list-item--orphan"' if item["orphan"] else "" + rendered.append(f"{content}{''.join(str(value) for value in nested)}") + return "".join(rendered) + + +def _collect_title_anchor_ids(middle_json: MiddleJson) -> dict[str, str]: + """为正文首次出现的非空 anchor 分配无空白且 document-wide 唯一的 HTML id。""" + anchor_ids: dict[str, str] = {} + used_ids: set[str] = set() + for page in middle_json.pages: + for block in page.blocks: + if not isinstance(block, TitleBlockBase): + continue + anchor = _anchor_key(block.anchor) + visible_title = inline_plain_text(parse_inline_content(block.content)).strip() + if not anchor or not visible_title or anchor in anchor_ids: + continue + base_id = _html_anchor_base(anchor) + candidate = base_id + suffix = 2 + while candidate in used_ids: + candidate = f"{base_id}-{suffix}" + suffix += 1 + anchor_ids[anchor] = candidate + used_ids.add(candidate) + return anchor_ids + + +def _plain_content_text(content: str) -> str: + """从 body 内容提取安全 img alt 所需的可见纯文本。""" + if not content: + return "" + if is_supported_html_markup(content): + return BeautifulSoup(content, "html.parser").get_text(" ", strip=True) + return inline_plain_text(parse_inline_content(content)).strip() + + +def _extract_mermaid_flowchart_source(content: str) -> str | None: + """提取受限 Mermaid flowchart fence，拒绝其他图类型和可改写配置的语法。""" + normalized = _replace_html_controls(content).replace("\r\n", "\n").replace("\r", "\n").strip() + if not normalized: + return None + lines = normalized.split("\n") + if len(lines) < 3: + return None + opening = re.fullmatch(r"(?P\x60{3,})[ \t]*mermaid[ \t]*", lines[0], re.IGNORECASE) + if opening is None: + return None + fence_length = len(opening.group("fence")) + if re.fullmatch(rf"\x60{{{fence_length},}}[ \t]*", lines[-1]) is None: + return None + source = "\n".join(lines[1:-1]).strip("\n") + if not source or len(source) > _MERMAID_MAX_TEXT_SIZE: + return None + if source.lstrip().startswith("---") or "%%{" in source: + return None + first_statement = next( + (line.strip() for line in source.splitlines() if line.strip() and not line.lstrip().startswith("%%")), + "", + ) + if _MERMAID_FLOWCHART_HEADER_RE.match(first_statement) is None: + return None + return source + + +def _contains_usable_table(markup: str) -> bool: + """确认清洗后的 markup 至少保留一个含单元格的 table。""" + soup = BeautifulSoup(markup, "html.parser") + return any(table.find(("td", "th")) is not None for table in soup.find_all("table")) + + +def _render_raw_fallback(content: str) -> str: + """把无法安全结构化的原内容转义为可见 pre，避免静默丢失。""" + return f'
{html.escape(_replace_html_controls(content), quote=False)}
' + + +def _normalize_prism_language(language: str | None) -> str | None: + """把模型语言名归一化为不可构造路径穿越的 Prism component 名。""" + normalized = (language or "").strip().lower() + normalized = _PRISM_LANGUAGE_ALIASES.get(normalized, normalized.replace("_", "-")) + if not normalized or normalized in _BLOCKED_PRISM_LANGUAGE_NAMES: + return None + return normalized if _SAFE_PRISM_LANGUAGE_RE.fullmatch(normalized) else None + + +def _resolve_document_title(middle_json: MiddleJson, explicit_title: str | None) -> str: + """按显式参数、首个正文文档标题、固定回退值解析 HTML title。""" + if explicit_title and explicit_title.strip(): + return explicit_title.strip() + for page in middle_json.pages: + for block in page.blocks: + if isinstance(block, DocTitleBlock): + title = inline_plain_text(parse_inline_content(block.content)).strip() + if title: + return title + return "MinerU Document" + + +def _anchor_key(anchor: str | None) -> str: + """保留 producer anchor 身份，仅去除首尾空白供标题与目录精确匹配。""" + return (anchor or "").strip() + + +def _html_anchor_base(anchor: str) -> str: + """把原始 anchor 转成无空白的 HTML id 基值，碰撞由 document registry 处理。""" + return re.sub(r"\s+", "-", _replace_html_controls(anchor)) + + +def _replace_html_controls(content: str) -> str: + """替换 HTML 不允许的 C0 控制字符与不可编码的孤立 surrogate。""" + return re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f\ud800-\udfff]", "\ufffd", content) + + +@lru_cache(maxsize=1) +def _load_html_styles() -> str: + """从包资源读取并缓存压缩后的 MinerU 独立样式。""" + root = resources.files("mineru").joinpath("resources", "html") + return root.joinpath(_STYLE_RESOURCE_NAME).read_text(encoding="utf-8") + + +def _mathjax_head() -> str: + """返回固定、收紧公式输入能力的 MathJax 4.1.2 head 片段。""" + return f""" +""" + + +def _prism_head() -> str: + """返回固定 Prism core、Autoloader 和当前文档局部高亮脚本。""" + return f""" + +""" + + +def _mermaid_head() -> str: + """返回固定 Mermaid 入口和逐图安全渲染、失败回退脚本。""" + return f""" +""" + + +__all__ = ["render_html"] diff --git a/mineru/render/_internal/html/sanitizer.py b/mineru/render/_internal/html/sanitizer.py index 97fae212..50d098bb 100644 --- a/mineru/render/_internal/html/sanitizer.py +++ b/mineru/render/_internal/html/sanitizer.py @@ -10,7 +10,7 @@ from urllib.parse import quote, unquote, urlsplit import nh3 from bs4 import BeautifulSoup, NavigableString, Tag -from mineru.utils.image_payload import parse_image_data_uri_strict +from ....utils.image_payload import parse_image_data_uri_strict _ALLOWED_TAGS = { diff --git a/mineru/render/_internal/html/table.py b/mineru/render/_internal/html/table.py index 715afedc..46d04ec2 100644 --- a/mineru/render/_internal/html/table.py +++ b/mineru/render/_internal/html/table.py @@ -5,7 +5,7 @@ from __future__ import annotations import re -from mineru.render._internal.html.inline import HtmlInlineResult, render_inline_content_html +from .inline import HtmlInlineResult, render_inline_content_html _SEPARATOR_CELL_RE = re.compile(r"^:?-{3,}:?$") diff --git a/mineru/render/_internal/markdown/__init__.py b/mineru/render/_internal/markdown/__init__.py index fd246652..f32d854d 100644 --- a/mineru/render/_internal/markdown/__init__.py +++ b/mineru/render/_internal/markdown/__init__.py @@ -1 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. """Markdown 与 Content List 共用的内部实现。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/markdown/assets.py b/mineru/render/_internal/markdown/assets.py index a1d2fb3f..c6b8a015 100644 --- a/mineru/render/_internal/markdown/assets.py +++ b/mineru/render/_internal/markdown/assets.py @@ -6,7 +6,7 @@ from __future__ import annotations import re from urllib.parse import quote -from mineru.types import ImagePayloadBlock +from ....types import ImagePayloadBlock _HTML_IMAGE_SRC_RE = re.compile( r"(?P\bsrc\s*=\s*)(?P[\"'])(?P.*?)(?P=quote)", diff --git a/mineru/render/_internal/markdown/blocks.py b/mineru/render/_internal/markdown/blocks.py index a0c7663a..8b8885a3 100644 --- a/mineru/render/_internal/markdown/blocks.py +++ b/mineru/render/_internal/markdown/blocks.py @@ -6,13 +6,13 @@ from __future__ import annotations import html import re -from mineru.config import LatexDelimitersConfig -from mineru.render._internal.common.index import strip_index_page_tail -from mineru.render._internal.common.list_items import has_markdown_unordered_marker, reference_list_needs_bullets -from mineru.render._internal.common.planner import PlannedBlock -from mineru.render._internal.markdown.assets import build_markdown_image, resolve_image_source -from mineru.render._internal.markdown.escaping import escape_standalone_marker_rule, escape_text_block_markdown_prefix -from mineru.render._internal.markdown.inline import ( +from ....config import LatexDelimitersConfig +from ..common.index import strip_index_page_tail +from ..common.list_items import has_markdown_unordered_marker, reference_list_needs_bullets +from ..common.planner import PlannedBlock +from .assets import build_markdown_image, resolve_image_source +from .escaping import escape_standalone_marker_rule, escape_text_block_markdown_prefix +from .inline import ( render_inline_content, render_internal_link, render_joined_inline_contents, @@ -23,7 +23,7 @@ from .table import ( render_html_table, ) from ...contracts import ImageRenderer -from mineru.types import ( +from ....types import ( RAW_ALGORITHM, BlockType, ChartAnnotationBlock, diff --git a/mineru/render/_internal/markdown/inline.py b/mineru/render/_internal/markdown/inline.py index 7d58698f..48284217 100644 --- a/mineru/render/_internal/markdown/inline.py +++ b/mineru/render/_internal/markdown/inline.py @@ -6,8 +6,8 @@ from __future__ import annotations import html import re -from mineru.config import LatexDelimitersConfig -from mineru.render._internal.common.inline import ( +from ....config import LatexDelimitersConfig +from ....backend.postprocess.inline import ( InlineEquation, InlineLink, InlineNode, @@ -17,7 +17,7 @@ from mineru.render._internal.common.inline import ( join_inline_contents, parse_inline_content, ) -from mineru.render._internal.markdown.escaping import escape_conservative_markdown_text +from .escaping import escape_conservative_markdown_text _ANGLE_TEXT_RE = re.compile(r"<[^<>\n]*>") _SIMPLE_STYLE_WRAPPERS = { diff --git a/mineru/render/_internal/markdown/renderer.py b/mineru/render/_internal/markdown/renderer.py new file mode 100644 index 00000000..e621771c --- /dev/null +++ b/mineru/render/_internal/markdown/renderer.py @@ -0,0 +1,71 @@ +# Copyright (c) Opendatalab. All rights reserved. +"""严格 MiddleJson 到 Markdown 的公共渲染实现。""" + +from __future__ import annotations + +from ....config import LatexDelimitersConfig, config +from ....types import PAGE_AUXILIARY_BLOCK_TYPES, MiddleJson +from ...contracts import ImageRenderer, RenderMode +from ..common.planner import PlannedBlock, build_render_plan +from .blocks import render_planned_block + +_PAGE_SEPARATOR = "\n\n---\n\n" + + +def render_markdown( + middle_json: MiddleJson, + *, + mode: RenderMode = RenderMode.DEFAULT, + asset_base_url: str = "", + image_renderer: ImageRenderer | None = None, +) -> str: + """把严格 MiddleJson 纯函数式渲染为 Markdown 字符串。""" + if not isinstance(middle_json, MiddleJson): + raise TypeError("render_markdown expects a MiddleJson instance") + if not isinstance(mode, RenderMode): + raise TypeError("mode must be a RenderMode value") + + delimiters = config.render.latex_delimiters + planned_pages = build_render_plan(middle_json, mode) + rendered_pages = [ + _render_page( + page, + mode=mode, + delimiters=delimiters, + asset_base_url=asset_base_url, + image_renderer=image_renderer, + ) + for page in planned_pages + ] + if mode is RenderMode.FULL: + return _PAGE_SEPARATOR.join(rendered_pages) + return "\n\n".join(page for page in rendered_pages if page) + + +def _render_page( + planned_blocks: list[PlannedBlock], + *, + mode: RenderMode, + delimiters: LatexDelimitersConfig, + asset_base_url: str, + image_renderer: ImageRenderer | None = None, +) -> str: + """渲染单页逻辑块，并在默认模式中过滤重复页元素。""" + rendered: list[str] = [] + for planned in planned_blocks: + if planned.removed: + continue + if mode is RenderMode.DEFAULT and planned.block.type in PAGE_AUXILIARY_BLOCK_TYPES: + continue + text = render_planned_block( + planned, + delimiters=delimiters, + asset_base_url=asset_base_url, + image_renderer=image_renderer, + ) + if text and text.strip(): + rendered.append(text.strip("\n")) + return "\n\n".join(rendered) + + +__all__ = ["render_markdown"] diff --git a/mineru/render/_internal/structured_content/__init__.py b/mineru/render/_internal/structured_content/__init__.py new file mode 100644 index 00000000..699f4393 --- /dev/null +++ b/mineru/render/_internal/structured_content/__init__.py @@ -0,0 +1,4 @@ +# Copyright (c) Opendatalab. All rights reserved. +"""Structured Content 私有渲染实现。""" + +__all__: list[str] = [] diff --git a/mineru/render/_internal/structured_content/renderer.py b/mineru/render/_internal/structured_content/renderer.py new file mode 100644 index 00000000..0e57823d --- /dev/null +++ b/mineru/render/_internal/structured_content/renderer.py @@ -0,0 +1,181 @@ +# Copyright (c) Opendatalab. All rights reserved. +"""严格 MiddleJson 到树形 Markdown structured_content 的公共渲染实现。""" + +from __future__ import annotations + +from typing import Any, TypeAlias + +from ....config import LatexDelimitersConfig, config +from ..markdown.blocks import ( + render_single_block, + render_title_inline_content, + render_visual_annotation, + render_visual_body_content, +) +from ..markdown.assets import normalize_image_source, resolve_image_source +from ..markdown.escaping import escape_standalone_marker_rule, escape_text_block_markdown_prefix +from ....types import ( + BlockType, + ChartAnnotationBlock, + ChartBlock, + ChartBodyBlock, + CodeAnnotationBlock, + CodeBlock, + DocTitleBlock, + EquationBlock, + ImageAnnotationBlock, + ImageBlock, + ImageBodyBlock, + ImagePayloadBlock, + MiddleJson, + PageBlock, + ParagraphTitleBlock, + TableAnnotationBlock, + TableBlock, + TableBodyBlock, +) + +VisualBlock: TypeAlias = ImageBlock | TableBlock | ChartBlock | CodeBlock +VisualAnnotationBlock: TypeAlias = ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock + +_CAPTION_TYPES = { + BlockType.IMAGE_CAPTION, + BlockType.TABLE_CAPTION, + BlockType.CHART_CAPTION, + BlockType.CODE_CAPTION, +} +_FOOTNOTE_TYPES = { + BlockType.IMAGE_FOOTNOTE, + BlockType.TABLE_FOOTNOTE, + BlockType.CHART_FOOTNOTE, + BlockType.CODE_FOOTNOTE, +} +_REMOVED_BLOCK_FIELDS = {"content", "index", "guess_lang", "image_path", "image_base64"} + + +def render_structured_content( + middle_json: MiddleJson, + *, + asset_base_url: str = "", +) -> dict[str, Any]: + """把严格 MiddleJson 无副作用地渲染为树形 Markdown structured_content。""" + if not isinstance(middle_json, MiddleJson): + raise TypeError("render_structured_content expects a MiddleJson instance") + + delimiters = config.render.latex_delimiters + document_fields = middle_json.model_dump( + mode="json", + exclude={"pages"}, + exclude_defaults=True, + ) + pages = [ + { + "page_idx": page.page_idx, + "blocks": [ + _render_content_block( + block, + delimiters=delimiters, + asset_base_url=asset_base_url, + ) + for block in page.blocks + ], + } + for page in middle_json.pages + ] + return {"pages": pages, **document_fields} + + +def _render_content_block( + block: PageBlock, + *, + delimiters: LatexDelimitersConfig, + asset_base_url: str, +) -> dict[str, Any]: + """保留父块元数据，并把 block 内容收敛为 Markdown 字符串。""" + payload = block.model_dump( + mode="json", + exclude=_REMOVED_BLOCK_FIELDS, + exclude_defaults=True, + ) + if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): + content = render_title_inline_content(block, delimiters) + payload["content"] = escape_standalone_marker_rule(escape_text_block_markdown_prefix(content)) + return payload + if isinstance(block, EquationBlock): + payload["content"] = block.content.strip() + image_source = _resolve_content_image_source(block, asset_base_url) + if image_source is not None: + payload["image_source"] = image_source + return payload + if not isinstance(block, (ImageBlock, TableBlock, ChartBlock, CodeBlock)): + payload["content"] = render_single_block( + block, + delimiters=delimiters, + asset_base_url=asset_base_url, + ) + return payload + + payload["content"] = render_visual_body_content( + block, + delimiters=delimiters, + asset_base_url=asset_base_url, + ) + payload["captions"] = _render_annotation_group(block, _CAPTION_TYPES, delimiters) + payload["footnotes"] = _render_annotation_group(block, _FOOTNOTE_TYPES, delimiters) + image_source = _resolve_visual_image_source(block, asset_base_url) + if image_source is not None: + payload["image_source"] = image_source + return payload + + +def _render_annotation_group( + block: VisualBlock, + accepted_types: set[str], + delimiters: LatexDelimitersConfig, +) -> list[dict[str, Any]]: + """按源 index 稳定排序视觉说明，并保留可用 bbox 与 Markdown 内容。""" + annotations: list[tuple[int, VisualAnnotationBlock]] = [] + for position, child in enumerate(block.content): + if ( + isinstance( + child, + (ImageAnnotationBlock, TableAnnotationBlock, ChartAnnotationBlock, CodeAnnotationBlock), + ) + and child.type in accepted_types + ): + annotations.append((position, child)) + annotations.sort(key=_annotation_sort_key) + rendered_annotations: list[dict[str, Any]] = [] + for _, child in annotations: + annotation_payload: dict[str, Any] = {} + if child.bbox is not None: + annotation_payload["bbox"] = list(child.bbox) + annotation_payload["content"] = render_visual_annotation(child, delimiters) + rendered_annotations.append(annotation_payload) + return rendered_annotations + + +def _annotation_sort_key( + item: tuple[int, VisualAnnotationBlock], +) -> tuple[bool, int, int]: + """让有 index 的说明升序优先，缺失 index 的说明稳定排在末尾。""" + position, block = item + return block.index is None, block.index if block.index is not None else 0, position + + +def _resolve_visual_image_source(block: VisualBlock, asset_base_url: str) -> str | None: + """解析视觉 body 实际选择的图片来源，并返回安全的 Markdown 地址。""" + for child in block.content: + if not isinstance(child, (ImageBodyBlock, TableBodyBlock, ChartBodyBlock)): + continue + return _resolve_content_image_source(child, asset_base_url) + return None + + +def _resolve_content_image_source(block: ImagePayloadBlock, asset_base_url: str) -> str | None: + """把图片载荷收敛为 structured_content 中唯一且安全的 image_source。""" + source = resolve_image_source(block, asset_base_url) + return normalize_image_source(source) if source else None + + +__all__ = ["render_structured_content"] diff --git a/mineru/render/api.py b/mineru/render/api.py index e569de26..9a6bab77 100644 --- a/mineru/render/api.py +++ b/mineru/render/api.py @@ -5,7 +5,7 @@ from __future__ import annotations from typing import Any, Literal, overload -from mineru.types import MiddleJson +from ..types import MiddleJson from .contracts import ( DocxRenderOptions, diff --git a/mineru/render/docx.py b/mineru/render/docx.py index 414c183f..163f3075 100644 --- a/mineru/render/docx.py +++ b/mineru/render/docx.py @@ -1,110 +1,10 @@ # Copyright (c) Opendatalab. All rights reserved. -"""严格 MiddleJson 到可编辑 DOCX bytes 的公共渲染实现。""" +"""严格 MiddleJson 到 DOCX 的轻量公共门面与稳定异常。""" from __future__ import annotations -from collections.abc import Iterable -from io import BytesIO -import re - -from bs4 import BeautifulSoup, NavigableString, Tag -from docx import Document -from docx.document import Document as DocumentType -from docx.enum.table import WD_CELL_VERTICAL_ALIGNMENT -from docx.enum.text import WD_ALIGN_PARAGRAPH, WD_TAB_ALIGNMENT -from docx.oxml import OxmlElement -from docx.oxml.ns import qn -from docx.shared import Emu, Mm, Pt, Twips -from docx.table import _Cell -from docx.text.paragraph import Paragraph -from lxml import etree -from loguru import logger - -from mineru.render._internal.common.index import strip_index_page_tail -from mineru.render._internal.common.inline import ( - InlineEquation, - InlineLink, - InlineNode, - InlineStyled, - InlineText, - parse_inline_content, -) -from mineru.render._internal.common.list_items import parse_list_item_marker -from mineru.render._internal.common.planner import PlannedBlock, build_render_plan -from mineru.render._internal.docx.assets import ( - DocxAssetError, - PreparedImage, - prepare_block_image, - prepare_html_image, -) -from mineru.render._internal.docx.inline import ( - BookmarkRegistry, - InlineRenderContext, - append_inline_content, - append_inline_nodes, - append_internal_link, - append_joined_inline_contents, - sanitize_xml_text, -) -from mineru.render._internal.docx.math import DocxFormulaError, latex_to_omml, split_formula_tag -from mineru.render._internal.docx.styles import ( - AUXILIARY_STYLE, - BODY_STYLE, - CAPTION_STYLE, - CODE_STYLE, - FOOTNOTE_STYLE, - FORMULA_FALLBACK_STYLE, - SPATIAL_TABLE_STYLE, - configure_document, - usable_width_emu, - usable_width_twips, -) -from mineru.render._internal.docx.table import DocxTableError, NestedTableWriter, materialize_docx_tables -from mineru.render.contracts import AssetResolver, RenderMode -from mineru.types import ( - PAGE_AUXILIARY_BLOCK_TYPES, - RAW_ALGORITHM, - BlockBase, - BlockType, - BBox, - ChartAnnotationBlock, - ChartBlock, - ChartBodyBlock, - CodeAnnotationBlock, - CodeBlock, - CodeBodyBlock, - DocTitleBlock, - EquationBlock, - ImageAnnotationBlock, - ImageBlock, - ImageBodyBlock, - ImagePayloadBlock, - IndexBlock, - ListBlock, - MiddleJson, - PageAuxTextBlock, - ParagraphTitleBlock, - RefTextBlock, - TableAnnotationBlock, - TableBlock, - TableBodyBlock, - TextBlock, - TitleBlockBase, -) - -_HTML_TABLE_RE = re.compile(r" None: - """初始化 renderer，并预注册全部标题 anchor。""" - self.middle_json = middle_json - self.mode = mode - self.asset_resolver = asset_resolver - self.document = Document() - configure_document(self.document) - self.bookmarks = BookmarkRegistry(_iter_document_anchors(middle_json)) - self.usable_width_emu = usable_width_emu(self.document) - self.usable_width_twips = usable_width_twips(self.document) - - def render(self) -> bytes: - """执行逐页 visitor，并把 python-docx document 序列化为 bytes。""" - planned_pages = build_render_plan(self.middle_json, self.mode) - for page_position, planned_blocks in enumerate(planned_pages): - if self.mode is RenderMode.FULL and page_position > 0: - self.document.add_page_break() - for planned in planned_blocks: - if planned.removed: - continue - if self.mode is RenderMode.DEFAULT and planned.block.type in PAGE_AUXILIARY_BLOCK_TYPES: - continue - self._render_planned_block(planned) - - output = BytesIO() - self.document.save(output) - return output.getvalue() - - def _render_planned_block(self, planned: PlannedBlock) -> None: - """按严格 PageBlock 具体类型分派 Word 写入逻辑。""" - block = planned.block - context = self._context(planned.page_idx, block) - if isinstance(block, (TextBlock, RefTextBlock)): - paragraph = self.document.add_paragraph(style=BODY_STYLE) - append_joined_inline_contents( - paragraph, - planned.text_contents or [block.content], - context=context, - ) - return - if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): - self._render_title(block, context) - return - if isinstance(block, PageAuxTextBlock): - paragraph = self.document.add_paragraph(style=AUXILIARY_STYLE) - append_inline_content(paragraph, block.content, context=context) - return - if isinstance(block, EquationBlock): - self._render_equation(block, context) - return - if isinstance(block, ListBlock): - self._render_list(block, context, depth=0) - return - if isinstance(block, IndexBlock): - self._render_index(block, context, depth=0) - return - if isinstance(block, ImageBlock): - self._render_image_block(block, context) - return - if isinstance(block, TableBlock): - self._render_table_block(block, context) - return - if isinstance(block, ChartBlock): - self._render_chart_block(block, context) - return - if isinstance(block, CodeBlock): - self._render_code_block(block, context) - return - raise TypeError(f"Unsupported PageBlock type: {type(block).__name__}") - - def _context(self, page_idx: int, block: BlockBase) -> InlineRenderContext: - """为当前 block 构造行内渲染上下文。""" - return InlineRenderContext( - bookmarks=self.bookmarks, - page_idx=page_idx, - block_index=block.index, - block_type=block.type, - ) - - def _render_title( - self, - block: DocTitleBlock | ParagraphTitleBlock, - context: InlineRenderContext, - ) -> None: - """用 Heading 1–6 写严格标题，并在段落内容外添加 bookmark。""" - level = min(max(block.level, 1), 9) - paragraph = self.document.add_paragraph(style=f"Heading {level}") - append_inline_content(paragraph, block.content, context=context) - self.bookmarks.attach(paragraph, block.anchor) - - def _render_equation(self, block: EquationBlock, context: InlineRenderContext) -> None: - """优先写可编辑 OMML，失败时按图片、可见 LaTeX 的顺序回退。""" - content = block.content.strip() - if content: - formula, tag = split_formula_tag(content) - try: - math_element = latex_to_omml(formula, display=True) - self._append_display_formula(math_element, tag, context=context) - return - except DocxFormulaError as exc: - logger.warning("DOCX display formula fallback: {} ({})", exc, context.location()) - - if block.image_base64 is not None or block.image_path is not None: - try: - self._append_block_image(block, context=context, alt_text="formula") - return - except DocxRenderError: - if not content: - raise - logger.warning("DOCX formula image fallback failed; preserving LaTeX ({})", context.location()) - - if content: - paragraph = self.document.add_paragraph(style=FORMULA_FALLBACK_STYLE) - paragraph.add_run(sanitize_xml_text(content, context=context)) - return - raise self._render_error("Equation does not contain usable LaTeX or image", context) - - def _append_display_formula( - self, - math_element: etree._Element, - tag: str | None, - *, - context: InlineRenderContext, - ) -> None: - """把块公式居中写入；带 tag 时使用中心与右对齐 tab。""" - if tag is None: - properties = OxmlElement("m:oMathParaPr") - justification = OxmlElement("m:jc") - justification.set(qn("m:val"), "center") - properties.append(justification) - math_element.insert(0, properties) - paragraph = self.document.add_paragraph(style=BODY_STYLE) - paragraph.paragraph_format.space_before = Pt(4) - paragraph.paragraph_format.space_after = Pt(5) - paragraph._p.append(math_element) - return - - equation = math_element.find(qn("m:oMath")) - if equation is None: - raise ValueError("Display OMML does not contain m:oMath") - math_element.remove(equation) - paragraph = self.document.add_paragraph(style=BODY_STYLE) - paragraph.paragraph_format.space_before = Pt(4) - paragraph.paragraph_format.space_after = Pt(5) - paragraph.paragraph_format.tab_stops.add_tab_stop( - Twips(self.usable_width_twips // 2), - WD_TAB_ALIGNMENT.CENTER, - ) - paragraph.paragraph_format.tab_stops.add_tab_stop( - Twips(self.usable_width_twips), - WD_TAB_ALIGNMENT.RIGHT, - ) - paragraph.add_run().add_tab() - paragraph._p.append(equation) - paragraph.add_run().add_tab() - rendered_tag = tag if tag.startswith("(") and tag.endswith(")") else f"({tag})" - paragraph.add_run(sanitize_xml_text(rendered_tag, context=context)) - - def _render_list( - self, - block: ListBlock, - context: InlineRenderContext, - *, - depth: int, - ) -> None: - """保留叶子原 marker，并仅用缩进表达递归列表层级。""" - for child in block.content: - if isinstance(child, ListBlock): - self._render_list(child, context, depth=depth + 1) - continue - paragraph = self.document.add_paragraph(style=BODY_STYLE) - item = parse_list_item_marker(child.content) - if item.marker is None: - paragraph.paragraph_format.left_indent = Mm(depth * 6) - else: - paragraph.paragraph_format.left_indent = Mm((depth + 1) * 6) - paragraph.paragraph_format.first_line_indent = Mm(-6) - append_inline_content(paragraph, child.content, context=context) - - def _render_index( - self, - block: IndexBlock, - context: InlineRenderContext, - *, - depth: int, - ) -> None: - """递归输出目录项，并把标题叶子链接到预注册 bookmark。""" - for child in block.content: - if isinstance(child, IndexBlock): - self._render_index(child, context, depth=depth + 1) - continue - content = strip_index_page_tail(child.content) - nodes = parse_inline_content(content) - if not nodes: - continue - paragraph = self.document.add_paragraph(style=BODY_STYLE) - paragraph.paragraph_format.left_indent = Mm((depth + 1) * 6) - paragraph.paragraph_format.first_line_indent = Mm(-4) - paragraph.add_run("• ") - anchor = child.anchor if isinstance(child, TitleBlockBase) else None - append_internal_link(paragraph, nodes, anchor=anchor, context=context) - - def _render_image_block(self, block: ImageBlock, context: InlineRenderContext) -> None: - """按原始子块顺序写图片主体、caption 与 footnote。""" - for child in block.content: - if isinstance(child, ImageBodyBlock): - alt_text = _plain_html_text(child.content) or block.sub_type or "image" - self._append_block_image(child, context=context, alt_text=alt_text) - elif isinstance(child, ImageAnnotationBlock): - self._render_annotation(child, context) - else: - raise TypeError(f"Unsupported image child: {type(child).__name__}") - - def _render_table_block(self, block: TableBlock, context: InlineRenderContext) -> None: - """按原始子块顺序写原生 HTML table 或预格式空间投影文本。""" - for child in block.content: - if isinstance(child, TableBodyBlock): - if _HTML_TABLE_RE.search(child.content): - try: - self._append_html_tables(child.content, context=context, depth=0) - except DocxRenderError as exc: - logger.warning("DOCX HTML table fallback: {}", exc) - if child.image_base64 is None and child.image_path is None: - raise self._render_error( - "HTML table cannot be materialized and has no image fallback", - context, - ) from exc - self._append_block_image(child, context=context, alt_text="table") - else: - if child.content and not child.content.isspace(): - paragraph = self.document.add_paragraph(style=SPATIAL_TABLE_STYLE) - paragraph.add_run(sanitize_xml_text(child.content, context=context)) - elif child.image_base64 is not None or child.image_path is not None: - self._append_block_image(child, context=context, alt_text="table") - else: - raise self._render_error( - "Spatial table does not contain text content or image", - context, - ) - elif isinstance(child, TableAnnotationBlock): - self._render_annotation(child, context) - else: - raise TypeError(f"Unsupported table child: {type(child).__name__}") - - def _render_chart_block(self, block: ChartBlock, context: InlineRenderContext) -> None: - """先写图表图片，再把 HTML 结构化数据追加为原生表格。""" - for child in block.content: - if isinstance(child, ChartBodyBlock): - has_image = child.image_base64 is not None or child.image_path is not None - if has_image: - self._append_block_image( - child, - context=context, - alt_text=_plain_html_text(child.content) or block.sub_type or "chart", - ) - if _HTML_TABLE_RE.search(child.content): - try: - self._append_html_tables(child.content, context=context, depth=0) - except DocxRenderError as exc: - if not has_image: - raise - logger.warning("DOCX chart data table omitted after image fallback: {}", exc) - elif child.content.strip() and not has_image: - paragraph = self.document.add_paragraph(style=BODY_STYLE) - append_inline_content(paragraph, child.content, context=context) - elif not has_image and not child.content.strip(): - raise self._render_error("Chart does not contain image or structured content", context) - elif isinstance(child, ChartAnnotationBlock): - self._render_annotation(child, context) - else: - raise TypeError(f"Unsupported chart child: {type(child).__name__}") - - def _render_code_block(self, block: CodeBlock, context: InlineRenderContext) -> None: - """按原始子块顺序写代码或算法主体及说明。""" - for child in block.content: - if isinstance(child, CodeBodyBlock): - paragraph = self.document.add_paragraph(style=CODE_STYLE) - if block.sub_type == BlockType.CODE: - paragraph.add_run(sanitize_xml_text(child.content, context=context)) - elif block.sub_type == RAW_ALGORITHM: - append_inline_content(paragraph, child.content, context=context) - else: - raise ValueError(f"Unsupported code subtype: {block.sub_type}") - elif isinstance(child, CodeAnnotationBlock): - self._render_annotation(child, context) - else: - raise TypeError(f"Unsupported code child: {type(child).__name__}") - - def _render_annotation( - self, - block: ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock, - context: InlineRenderContext, - ) -> None: - """根据说明类型选择 Caption 或 Footnote 样式并写入行内内容。""" - if block.type in _ANNOTATION_CAPTION_TYPES: - style = CAPTION_STYLE - elif block.type in _ANNOTATION_FOOTNOTE_TYPES: - style = FOOTNOTE_STYLE - else: - raise TypeError(f"Unsupported annotation type: {block.type}") - paragraph = self.document.add_paragraph(style=style) - append_inline_content(paragraph, block.content, context=context) - - def _append_block_image( - self, - block: ImagePayloadBlock, - *, - context: InlineRenderContext, - alt_text: str, - ) -> None: - """安全加载 block 图片，并按 bbox/自然尺寸限制到可用页面范围。""" - try: - prepared = prepare_block_image(block, self.asset_resolver) - except DocxAssetError as exc: - raise self._render_error(str(exc), context) from exc - self._append_prepared_image( - prepared, - bbox=block.bbox, - alt_text=alt_text, - max_width_emu=self.usable_width_emu, - context=context, - ) - - def _append_prepared_image( - self, - prepared: PreparedImage, - *, - bbox: BBox | None, - alt_text: str, - max_width_emu: int, - context: InlineRenderContext, - paragraph: Paragraph | None = None, - ) -> None: - """把准备好的图片按比例插入居中段落并写入 alt description。""" - target_paragraph = paragraph or self.document.add_paragraph(style=BODY_STYLE) - target_paragraph.alignment = WD_ALIGN_PARAGRAPH.CENTER - natural_width = round(prepared.width_px / 96 * 914400) - desired_width = natural_width - if bbox is not None: - desired_width = round(max_width_emu * max(0.0, min(1.0, bbox[2] - bbox[0]))) - minimum_width = min(int(Mm(5)), max_width_emu) - desired_width = max(minimum_width, min(desired_width, max_width_emu)) - - usable_height = int(Mm(257)) - rendered_height = round(desired_width * prepared.height_px / max(prepared.width_px, 1)) - if rendered_height > usable_height: - desired_width = round(usable_height * prepared.width_px / max(prepared.height_px, 1)) - - shape = target_paragraph.add_run().add_picture( - BytesIO(prepared.data), - width=Emu(max(1, desired_width)), - ) - description = sanitize_xml_text( - re.sub(r"\s+", " ", alt_text).strip()[:2048], - context=context, - ) - if description: - shape._inline.docPr.set("descr", description) - - def _append_html_tables( - self, - markup: str, - *, - context: InlineRenderContext, - depth: int, - container: DocumentType | _Cell | None = None, - ) -> None: - """物化 HTML 表格，并通过共享行内/素材逻辑填充 origin cell。""" - if depth >= 4: - raise self._render_error("Nested HTML table depth exceeds 4", context) - target_container = container or self.document - target_xml = target_container._element.body if isinstance(target_container, DocumentType) else target_container._tc - existing_child_count = len(target_xml) - insertion_index = existing_child_count - 1 if isinstance(target_container, DocumentType) else existing_child_count - existing_relationship_ids = set(self.document.part.rels) - - def fill_cell(cell: _Cell, source: Tag, write_nested: NestedTableWriter) -> None: - """填充一个 origin cell，并在遇到 nested table 时调用递归 writer。""" - self._fill_html_cell( - cell, - source, - write_nested=write_nested, - context=context, - ) - - try: - materialize_docx_tables( - target_container, - markup, - width_twips=self.usable_width_twips, - fill_cell=fill_cell, - ) - except DocxRenderError: - self._rollback_table_materialization( - target_xml, - existing_child_count=existing_child_count, - insertion_index=insertion_index, - existing_relationship_ids=existing_relationship_ids, - ) - raise - except DocxTableError as exc: - self._rollback_table_materialization( - target_xml, - existing_child_count=existing_child_count, - insertion_index=insertion_index, - existing_relationship_ids=existing_relationship_ids, - ) - raise self._render_error(str(exc), context) from exc - - def _rollback_table_materialization( - self, - target_xml: etree._Element, - *, - existing_child_count: int, - insertion_index: int, - existing_relationship_ids: set[str], - ) -> None: - """回滚本轮表格新增的 XML 根节点和 relationship，避免残留半成品。""" - while len(target_xml) > existing_child_count: - target_xml.remove(target_xml[insertion_index]) - for relationship_id in tuple(self.document.part.rels): - if relationship_id not in existing_relationship_ids: - self.document.part.drop_rel(relationship_id) - - def _fill_html_cell( - self, - cell: _Cell, - source: Tag, - *, - write_nested: NestedTableWriter, - context: InlineRenderContext, - ) -> None: - """按源顺序填充单元格文本、图片和任意包装层中的嵌套表格。""" - cell.text = "" - cell.vertical_alignment = WD_CELL_VERTICAL_ALIGNMENT.CENTER - _set_cell_margins(cell, top=100, start=120, bottom=100, end=120) - max_image_width_emu = _cell_content_width_emu(cell, horizontal_margin_twips=240) - paragraph = cell.paragraphs[0] - paragraph.paragraph_format.space_after = Pt(0) - paragraph.paragraph_format.line_spacing = 1.0 - - for child in source.children: - if not isinstance(child, (NavigableString, Tag)): - continue - if isinstance(child, Tag) and child.name.lower() in {"p", "div"} and _paragraph_has_content(paragraph): - paragraph = cell.add_paragraph() - paragraph.paragraph_format.space_after = Pt(0) - paragraph.paragraph_format.line_spacing = 1.0 - styles = ("bold",) if source.name == "th" else () - paragraph, _ = self._append_html_node( - paragraph, - child, - cell=cell, - write_nested=write_nested, - context=context, - inherited_styles=styles, - max_image_width_emu=max_image_width_emu, - ) - - def _append_html_node( - self, - paragraph: Paragraph, - node: NavigableString | Tag, - *, - cell: _Cell, - write_nested: NestedTableWriter, - context: InlineRenderContext, - inherited_styles: tuple[str, ...], - max_image_width_emu: int, - ) -> tuple[Paragraph, bool]: - """按 HTML 源顺序写入单元格行内文本、样式、公式和图片。""" - if isinstance(node, NavigableString): - text = str(node) - if not text: - return paragraph, False - nodes: list[InlineNode] = [InlineText(text)] - if inherited_styles: - nodes = [InlineStyled(nodes, inherited_styles)] - append_inline_nodes(paragraph, nodes, context=context) - return paragraph, bool(text.strip()) - - name = node.name.lower() - if name == "table": - write_nested(node) - trailing_paragraph = cell.paragraphs[-1] - trailing_paragraph.paragraph_format.space_after = Pt(0) - trailing_paragraph.paragraph_format.line_spacing = 1.0 - return trailing_paragraph, True - if name == "img": - self._append_html_image( - paragraph, - str(node.get("src", "")), - context=context, - alt_text=str(node.get("alt", "")), - max_width_emu=max_image_width_emu, - ) - return paragraph, True - if name == "br": - paragraph.add_run().add_break() - return paragraph, True - if name == "eq": - latex = node.get_text().strip() - nodes = [InlineEquation(latex)] if latex else [] - if inherited_styles and nodes: - nodes = [InlineStyled(nodes, inherited_styles)] - append_inline_nodes(paragraph, nodes, context=context) - return paragraph, bool(nodes) - if name == "a": - children = _html_inline_nodes(node) - if inherited_styles and children: - children = [InlineStyled(children, inherited_styles)] - append_inline_nodes( - paragraph, - [InlineLink(children, str(node.get("href", "")).strip())], - context=context, - ) - return paragraph, bool(children) - if name in {"ul", "ol"}: - start_value = node.get("start", "1") - try: - ordered_number = int(str(start_value)) - except ValueError: - ordered_number = 1 - rendered = False - for item in node.find_all("li", recursive=False): - if _paragraph_has_content(paragraph) or rendered: - paragraph.add_run().add_break() - marker = f"{ordered_number}. " if name == "ol" else "- " - paragraph.add_run(marker) - if name == "ol": - ordered_number += 1 - for child in item.children: - if isinstance(child, (NavigableString, Tag)): - paragraph, child_rendered = self._append_html_node( - paragraph, - child, - cell=cell, - write_nested=write_nested, - context=context, - inherited_styles=inherited_styles, - max_image_width_emu=max_image_width_emu, - ) - rendered = child_rendered or rendered - return paragraph, rendered - if name == "li": - rendered = False - for child in node.children: - if isinstance(child, (NavigableString, Tag)): - paragraph, child_rendered = self._append_html_node( - paragraph, - child, - cell=cell, - write_nested=write_nested, - context=context, - inherited_styles=inherited_styles, - max_image_width_emu=max_image_width_emu, - ) - rendered = child_rendered or rendered - return paragraph, rendered - if name == "text": - nodes = parse_inline_content(str(node)) - if inherited_styles and nodes: - nodes = [InlineStyled(nodes, inherited_styles)] - append_inline_nodes(paragraph, nodes, context=context) - return paragraph, bool(nodes) - - style = { - "strong": "bold", - "b": "bold", - "em": "italic", - "i": "italic", - "u": "underline", - "s": "strikethrough", - "sup": "superscript", - "sub": "subscript", - }.get(name) - styles = tuple(dict.fromkeys((*inherited_styles, style))) if style else inherited_styles - rendered = False - for child in node.children: - if isinstance(child, (NavigableString, Tag)): - paragraph, child_rendered = self._append_html_node( - paragraph, - child, - cell=cell, - write_nested=write_nested, - context=context, - inherited_styles=styles, - max_image_width_emu=max_image_width_emu, - ) - rendered = child_rendered or rendered - return paragraph, rendered - - def _append_html_image( - self, - paragraph: Paragraph, - source: str, - *, - context: InlineRenderContext, - alt_text: str, - max_width_emu: int, - ) -> None: - """安全加载表格单元格 img，并限制到紧凑的单元格宽度。""" - try: - prepared = prepare_html_image(source, self.asset_resolver) - except DocxAssetError as exc: - raise self._render_error(str(exc), context) from exc - self._append_prepared_image( - prepared, - bbox=None, - alt_text=alt_text or "table image", - max_width_emu=min(self.usable_width_emu, max_width_emu), - context=context, - paragraph=paragraph, - ) - - def _render_error(self, message: str, context: InlineRenderContext) -> DocxRenderError: - """用当前 context 构造带 page/block 定位的公共异常。""" - return DocxRenderError( - message, - page_idx=context.page_idx, - block_index=context.block_index, - block_type=context.block_type, - ) - - def render_docx( middle_json: MiddleJson, *, mode: RenderMode = RenderMode.DEFAULT, asset_resolver: AssetResolver | None = None, ) -> bytes: - """把严格 MiddleJson 无副作用地渲染为完整 DOCX bytes。""" - if not isinstance(middle_json, MiddleJson): - raise TypeError("render_docx expects a MiddleJson instance") - if not isinstance(mode, RenderMode): - raise TypeError("mode must be a RenderMode value") - return _DocxRenderer( - middle_json, - mode=mode, - asset_resolver=asset_resolver, - ).render() + """惰性加载 DOCX 实现并渲染严格 MiddleJson。""" + from ._internal.docx.renderer import render_docx as _render_docx - -def _iter_document_anchors(middle_json: MiddleJson) -> Iterable[str]: - """按文档顺序遍历实际会写入 bookmark 的正文标题 anchor。""" - for page in middle_json.pages: - for block in page.blocks: - if isinstance(block, TitleBlockBase) and block.anchor: - yield block.anchor - - -def _plain_html_text(content: str) -> str: - """把图片或图表 body 内容收敛为适合 alt description 的纯文本。""" - if not content.strip(): - return "" - return BeautifulSoup(content, "html.parser").get_text(" ", strip=True) - - -def _html_inline_nodes(node: Tag) -> list[InlineNode]: - """把表格单元格中的安全 HTML 行内标签转换为共享 InlineNode。""" - nodes: list[InlineNode] = [] - for child in node.children: - if isinstance(child, NavigableString): - if str(child): - nodes.append(InlineText(str(child))) - continue - if not isinstance(child, Tag): - continue - name = child.name.lower() - children = _html_inline_nodes(child) - if name == "eq": - latex = child.get_text().strip() - if latex: - nodes.append(InlineEquation(latex)) - elif name in {"strong", "b"}: - nodes.append(InlineStyled(children, ("bold",))) - elif name in {"em", "i"}: - nodes.append(InlineStyled(children, ("italic",))) - elif name == "u": - nodes.append(InlineStyled(children, ("underline",))) - elif name == "s": - nodes.append(InlineStyled(children, ("strikethrough",))) - elif name == "sup": - nodes.append(InlineStyled(children, ("superscript",))) - elif name == "sub": - nodes.append(InlineStyled(children, ("subscript",))) - elif name == "a": - nodes.append(InlineLink(children, str(child.get("href", "")).strip())) - elif name == "br": - nodes.append(InlineText("\n")) - elif name == "text": - nodes.extend(parse_inline_content(str(child))) - elif name not in {"img", "table"}: - nodes.extend(children) - return nodes - - -def _set_cell_margins( - cell: _Cell, - *, - top: int, - start: int, - bottom: int, - end: int, -) -> None: - """用 DXA 为 Word 单元格写入四边内边距。""" - cell_properties = cell._tc.get_or_add_tcPr() - margins = cell_properties.find(qn("w:tcMar")) - if margins is None: - margins = OxmlElement("w:tcMar") - cell_properties.append(margins) - for side, value in (("top", top), ("start", start), ("bottom", bottom), ("end", end)): - element = margins.find(qn(f"w:{side}")) - if element is None: - element = OxmlElement(f"w:{side}") - margins.append(element) - element.set(qn("w:w"), str(value)) - element.set(qn("w:type"), "dxa") - - -def _cell_content_width_emu(cell: _Cell, *, horizontal_margin_twips: int) -> int: - """从当前物理或合并单元格 tcW 计算扣除左右内边距后的可用宽度。""" - cell_width = cell._tc.get_or_add_tcPr().get_or_add_tcW().w - width_twips = int(cell_width) if cell_width is not None else 1 - return int(Twips(max(1, width_twips - horizontal_margin_twips))) - - -def _paragraph_has_content(paragraph: Paragraph) -> bool: - """判断段落是否包含 pPr 之外的 run、公式、链接或 drawing。""" - return any(child.tag != qn("w:pPr") for child in paragraph._p) + return _render_docx(middle_json, mode=mode, asset_resolver=asset_resolver) __all__ = ["DocxRenderError", "render_docx"] diff --git a/mineru/render/html.py b/mineru/render/html.py index d906d6a9..94b34ebb 100644 --- a/mineru/render/html.py +++ b/mineru/render/html.py @@ -1,633 +1,10 @@ # Copyright (c) Opendatalab. All rights reserved. -"""严格 MiddleJson 到 MinerU 风格 HTML 的公共渲染实现。""" +"""严格 MiddleJson 到 HTML 的轻量公共门面。""" from __future__ import annotations -import html -import re -from functools import lru_cache -from importlib import resources -from urllib.parse import quote - -from bs4 import BeautifulSoup -from loguru import logger - -from mineru.render._internal.common.index import strip_index_page_tail -from mineru.render._internal.common.inline import inline_plain_text, parse_inline_content -from mineru.render._internal.common.list_items import ListItem, parse_list_item_marker, reference_list_needs_bullets -from mineru.render._internal.common.planner import PlannedBlock, build_render_plan -from mineru.render._internal.html.inline import ( - HtmlInlineResult, - render_inline_content_html, - render_inline_nodes_html, - render_joined_inline_contents_html, - render_math_html, -) -from mineru.render._internal.html.sanitizer import ( - is_supported_html_markup, - sanitize_html_fragment, - sanitize_image_source, -) -from mineru.render._internal.html.table import looks_like_gfm_table, render_gfm_table_html -from mineru.render.contracts import RenderMode -from mineru.types import ( - PAGE_AUXILIARY_BLOCK_TYPES, - RAW_ALGORITHM, - BlockType, - ChartAnnotationBlock, - ChartBlock, - ChartBodyBlock, - CodeAnnotationBlock, - CodeBlock, - CodeBodyBlock, - DocTitleBlock, - EquationBlock, - ImageAnnotationBlock, - ImageBlock, - ImageBodyBlock, - IndexBlock, - ListBlock, - MiddleJson, - PageAuxTextBlock, - ParagraphTitleBlock, - RefTextBlock, - TableAnnotationBlock, - TableBlock, - TableBodyBlock, - TextBlock, - TitleBlockBase, -) - -_STYLE_RESOURCE_NAME = "mineru.min.css" -_MATHJAX_URL = "https://cdn.jsdelivr.net/npm/mathjax@4.1.2/tex-chtml.js" -_MATHJAX_INTEGRITY = "sha384-zAhQQhdaMeHsMProNntGGg6nOUVcfuF9F22C3d1qJ9NZAVzCplXk1X85D2O5iufn" -_PRISM_CORE_URL = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/components/prism-core.min.js" -_PRISM_CORE_INTEGRITY = "sha384-zLRFO4dwowZvh8kzutOb5AWhH7f39HeJp+N7PtHF1SQtTBnifRx0AtmvTYs3F4YV" -_PRISM_AUTOLOADER_URL = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/plugins/autoloader/prism-autoloader.min.js" -_PRISM_AUTOLOADER_INTEGRITY = "sha384-Uq05+JLko69eOiPr39ta9bh7kld5PKZoU+fF7g0EXTAriEollhZ+DrN8Q/Oi8J2Q" -_PRISM_LANGUAGES_PATH = "https://cdn.jsdelivr.net/npm/prismjs@1.30.0/components/" -_MERMAID_URL = "https://cdn.jsdelivr.net/npm/mermaid@11.16.1/dist/mermaid.min.js" -_MERMAID_INTEGRITY = "sha384-aBQXj4hK6Jm05i7aQAsUV3bLdSUrHX1BGYfMB0166TtWt/RRaw+h0Eelme9OCOvy" -_MERMAID_MAX_TEXT_SIZE = 50_000 -_MERMAID_FLOWCHART_HEADER_RE = re.compile( - r"(?:graph|flowchart)\s+(?:TB|TD|BT|RL|LR)\b", - re.IGNORECASE, -) -_SAFE_PRISM_LANGUAGE_RE = re.compile(r"[a-z0-9][a-z0-9-]{0,31}") -_BLOCKED_PRISM_LANGUAGE_NAMES = {"constructor", "prototype"} -_PRISM_LANGUAGE_ALIASES = { - "c#": "csharp", - "c++": "cpp", - "html": "markup", - "js": "javascript", - "py": "python", - "sh": "bash", - "shell": "bash", - "none": "", - "text": "", - "txt": "", - "ts": "typescript", - "xml": "markup", - "yml": "yaml", -} - - -class _HtmlRenderer: - """维护一次 HTML 渲染所需的 anchor、公式与代码资源状态。""" - - def __init__( - self, - middle_json: MiddleJson, - *, - mode: RenderMode, - asset_base_url: str, - standalone: bool, - document_title: str | None, - ) -> None: - """保存严格输入和调用选项，并预收集真实正文标题 anchor。""" - self.middle_json = middle_json - self.mode = mode - self.asset_base_url = asset_base_url - self.standalone = standalone - self.document_title = document_title - self.anchor_targets = _collect_title_anchor_ids(middle_json) - self.emitted_anchors: set[str] = set() - self.has_math = False - self.has_prism = False - self.has_mermaid = False - - def render(self) -> str: - """生成 fragment article，并按需包装完整 HTML 文档。""" - planned_pages = build_render_plan(self.middle_json, self.mode) - if self.mode is RenderMode.FULL: - body = self._render_full_pages(planned_pages) - else: - body = self._render_default_pages(planned_pages) - article = f'
\n{body}\n
' - if not self.standalone: - return article - # 最终依赖以真实 DOM 为准，避免回退分支或普通文本中的 class 字样误触发。 - article_soup = BeautifulSoup(article, "html.parser") - self.has_math = article_soup.select_one(".mineru-math") is not None - self.has_prism = any( - any(str(class_name).startswith("language-") for class_name in code.get("class", [])) - for code in article_soup.find_all("code") - if code.get_text() - ) - self.has_mermaid = article_soup.select_one(".mineru-flowchart") is not None - return self._render_standalone(article) - - def _render_default_pages(self, planned_pages: list[list[PlannedBlock]]) -> str: - """把 DEFAULT 计划展平为无页面边界的连续阅读内容。""" - rendered: list[str] = [] - for page in planned_pages: - for planned in page: - block = self._render_planned_block(planned) - if block: - rendered.append(block) - return "\n".join(rendered) - - def _render_full_pages(self, planned_pages: list[list[PlannedBlock]]) -> str: - """为 FULL 计划保留所有页面 section 和相邻页面分隔。""" - sections: list[str] = [] - for page in planned_pages: - page_idx = page[0].page_idx if page else None - if page_idx is None: - page_position = len(sections) - page_idx = self.middle_json.pages[page_position].page_idx - blocks = [rendered for planned in page if (rendered := self._render_planned_block(planned))] - rendered_blocks = "\n".join(blocks) - sections.append(f'
\n{rendered_blocks}\n
') - return '\n\n'.join(sections) - - def _render_planned_block(self, planned: PlannedBlock) -> str: - """过滤计划块、分派类型 visitor，并添加稳定来源元数据。""" - if planned.removed: - return "" - block = planned.block - if self.mode is RenderMode.DEFAULT and block.type in PAGE_AUXILIARY_BLOCK_TYPES: - return "" - content = self._render_block_content(planned) - if not content or not content.strip(): - return "" - attrs = [ - 'class="mineru-block"', - f'data-page-idx="{planned.page_idx}"', - f'data-block-type="{html.escape(str(block.type), quote=True)}"', - ] - if block.index is not None: - attrs.append(f'data-block-index="{block.index}"') - return f"
\n{content}\n
" - - def _render_block_content(self, planned: PlannedBlock) -> str: - """按严格 PageBlock 具体类型返回对应语义 HTML。""" - block = planned.block - if isinstance(block, TextBlock): - rendered = render_joined_inline_contents_html(planned.text_contents or [block.content]) - self._observe_inline(rendered) - return f'

{rendered.html}

' if rendered.html else "" - if isinstance(block, RefTextBlock): - rendered = render_joined_inline_contents_html(planned.text_contents or [block.content]) - self._observe_inline(rendered) - return f'

{rendered.html}

' if rendered.html else "" - if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): - return self._render_title(block) - if isinstance(block, PageAuxTextBlock): - rendered = render_inline_content_html(block.content) - self._observe_inline(rendered) - class_name = html.escape(str(block.type).replace("_", "-"), quote=True) - return f'
{rendered.html}
' if rendered.html else "" - if isinstance(block, EquationBlock): - return self._render_equation(block) - if isinstance(block, ListBlock): - return self._render_list(block) - if isinstance(block, IndexBlock): - return self._render_index(block) - if isinstance(block, ImageBlock): - return self._render_image_block(block) - if isinstance(block, TableBlock): - return self._render_table_block(block) - if isinstance(block, ChartBlock): - return self._render_chart_block(block) - if isinstance(block, CodeBlock): - return self._render_code_block(block) - raise TypeError(f"Unsupported PageBlock type: {type(block).__name__}") - - def _render_title(self, block: DocTitleBlock | ParagraphTitleBlock) -> str: - """渲染标题，并保证正文 anchor 只在首次出现时生成 id。""" - rendered = render_inline_content_html(block.content) - self._observe_inline(rendered) - if not rendered.html: - return "" - level = min(max(block.level, 1), 6) - attrs = [f'class="mineru-heading mineru-heading--{level}"', f'data-heading-level="{block.level}"'] - anchor = _anchor_key(block.anchor) - if anchor: - if anchor not in self.emitted_anchors: - anchor_id = self.anchor_targets.get(anchor) - if anchor_id: - attrs.append(f'id="{html.escape(anchor_id, quote=True)}"') - self.emitted_anchors.add(anchor) - else: - logger.warning("Duplicate HTML title anchor omitted: {!r}", anchor) - return f"{rendered.html}" - - def _render_equation(self, block: EquationBlock) -> str: - """优先输出行间公式，空 LaTeX 时回退到安全图片来源。""" - if block.content.strip(): - rendered = render_math_html(block.content, display=True) - self._observe_inline(rendered) - return rendered.html - source = self._safe_block_image_source(block) - return self._render_image(source, alt="formula", class_name="mineru-equation-image") if source else "" - - def _render_list(self, block: ListBlock) -> str: - """按直属 marker 类型渲染一层列表，并递归挂接嵌套列表。""" - parsed_leaves = [ - parse_list_item_marker(child.content) - for child in block.content - if not isinstance(child, ListBlock) and child.content.strip() - ] - add_reference_bullets = reference_list_needs_bullets(block) - container_tag, list_type, class_name = _classify_list(parsed_leaves, add_reference_bullets) - items: list[dict[str, object]] = [] - expected_value: int | None = None - - for child in block.content: - if isinstance(child, ListBlock): - nested = self._render_list(child) - if not nested: - continue - if not items: - items.append({"content": "", "attrs": ' class="mineru-list-item--orphan"', "nested": []}) - nested_items = items[-1]["nested"] - assert isinstance(nested_items, list) - nested_items.append(nested) - continue - - parsed = parse_list_item_marker(child.content) - item_content, marker = _list_item_content( - parsed, - add_reference_bullets, - explicit_markers=class_name == "mineru-list--explicit", - ) - rendered = render_inline_content_html(item_content) - self._observe_inline(rendered) - if not rendered.html and not marker: - items.append({"content": "", "attrs": ' class="mineru-list-item--markerless"', "nested": []}) - continue - - attrs: list[str] = [] - if class_name == "mineru-list--explicit": - attrs.append('class="mineru-list-item--explicit"') - if container_tag == "ol" and parsed.kind == "ordered" and parsed.value is not None: - if expected_value is None: - expected_value = parsed.value - if parsed.value != expected_value: - attrs.append(f'value="{parsed.value}"') - expected_value = parsed.value + 1 - marker_html = ( - f'{html.escape(_replace_html_controls(marker), quote=False)}' - if marker - else "" - ) - if class_name == "mineru-list--explicit" and not marker_html: - marker_html = '' - items.append( - { - "content": f'{marker_html}{rendered.html}', - "attrs": f" {' '.join(attrs)}" if attrs else "", - "nested": [], - } - ) - - rendered_items: list[str] = [] - for item in items: - nested = item["nested"] - assert isinstance(nested, list) - item_content = str(item["content"]) - if not item_content and not nested: - continue - rendered_items.append(f"{item_content}{''.join(str(value) for value in nested)}") - if not rendered_items: - return "" - - container_attrs = [f'class="mineru-list {class_name}"'] - if list_type: - container_attrs.append(f'type="{list_type}"') - if container_tag == "ol" and parsed_leaves and parsed_leaves[0].value not in (None, 1): - container_attrs.append(f'start="{parsed_leaves[0].value}"') - return f"<{container_tag} {' '.join(container_attrs)}>{''.join(rendered_items)}" - - def _render_index(self, block: IndexBlock) -> str: - """递归渲染目录，并仅链接到真实正文标题 anchor。""" - items: list[dict[str, object]] = [] - for child in block.content: - if isinstance(child, IndexBlock): - nested = self._render_index_list(child) - if not nested: - continue - if not items: - items.append({"content": "", "orphan": True, "nested": []}) - nested_items = items[-1]["nested"] - assert isinstance(nested_items, list) - nested_items.append(nested) - continue - rendered = self._render_index_leaf(child) - items.append({"content": rendered, "orphan": False, "nested": []}) - inner = _serialize_index_items(items) - return f'' if inner else "" - - def _render_index_list(self, block: IndexBlock) -> str: - """渲染一个嵌套 IndexBlock 为可挂接到父项的 ul。""" - items: list[dict[str, object]] = [] - for child in block.content: - if isinstance(child, IndexBlock): - nested = self._render_index_list(child) - if nested: - if not items: - items.append({"content": "", "orphan": True, "nested": []}) - nested_items = items[-1]["nested"] - assert isinstance(nested_items, list) - nested_items.append(nested) - continue - rendered = self._render_index_leaf(child) - items.append({"content": rendered, "orphan": False, "nested": []}) - inner = _serialize_index_items(items) - return f"
    {inner}
" if inner else "" - - def _render_index_leaf(self, block: TextBlock | TitleBlockBase) -> str: - """渲染目录叶子，并在 anchor 命中正文标题时生成内部链接。""" - content = strip_index_page_tail(block.content) - rendered = render_inline_content_html(content) - self._observe_inline(rendered) - if not rendered.html: - return "" - anchor = _anchor_key(block.anchor) if isinstance(block, TitleBlockBase) else "" - if anchor and (anchor_id := self.anchor_targets.get(anchor)): - href = quote(anchor_id, safe="-._~") - return f'{rendered.html}' - return rendered.html - - def _render_image_block(self, block: ImageBlock) -> str: - """按原始子块顺序渲染图片主体及重复说明。""" - parts: list[str] = [] - for child in block.content: - if isinstance(child, ImageBodyBlock): - body = self._render_image_body(block, child) - elif isinstance(child, ImageAnnotationBlock): - body = self._render_annotation(child) - else: - raise TypeError(f"Unsupported image child: {type(child).__name__}") - if body: - parts.append(body) - return f'
{"".join(parts)}
' if parts else "" - - def _render_image_body(self, parent: ImageBlock, block: ImageBodyBlock) -> str: - """渲染图片，并把并存的识别内容放入原生 details。""" - if parent.sub_type == "flowchart": - mermaid_source = _extract_mermaid_flowchart_source(block.content) - if mermaid_source is not None: - return self._render_flowchart_body(block, mermaid_source) - source = self._safe_block_image_source(block) - content = self._render_embedded_content(block.content, linkify_text=parent.sub_type != "flowchart") - alt = _plain_content_text(block.content) or parent.sub_type or "image" - parts = [self._render_image(source, alt=alt, class_name="mineru-image")] if source else [] - if content.html: - if source: - parts.append(self._render_details(content.html, parent.sub_type or "image content")) - else: - parts.append(content.html) - return "".join(parts) - - def _render_flowchart_body(self, block: ImageBodyBlock, mermaid_source: str) -> str: - """输出 Mermaid canvas，并保留 raster 与源码两级失败回退。""" - source = self._safe_block_image_source(block) - escaped_source = html.escape(_replace_html_controls(mermaid_source), quote=False) - fallback = self._render_image(source, alt="flowchart", class_name="mineru-flowchart-fallback") if source else "" - fallback_class = " mineru-flowchart--has-raster" if fallback else "" - details_open = "" if fallback else " open" - return ( - f'
' - '' - f"{fallback}
" - f'
' - 'flowchart source' - f'
{escaped_source}
' - "
" - ) - - def _render_table_block(self, block: TableBlock) -> str: - """按原始子块顺序渲染表格主体、标题与脚注。""" - parts: list[str] = [] - for child in block.content: - if isinstance(child, TableBodyBlock): - body = self._render_table_body(child) - elif isinstance(child, TableAnnotationBlock): - body = self._render_annotation(child) - else: - raise TypeError(f"Unsupported table child: {type(child).__name__}") - if body: - parts.append(body) - return f'
{"".join(parts)}
' if parts else "" - - def _render_table_body(self, block: TableBodyBlock) -> str: - """优先保留安全 HTML table，空间文本和整体图片依次回退。""" - normalized_content = _replace_html_controls(block.content) - content = normalized_content.strip() - if content: - source_soup = BeautifulSoup(content, "html.parser") - if source_soup.find("table") is not None: - rendered = self._render_embedded_content(content) - if rendered.html and _contains_usable_table(rendered.html): - return rendered.html - source = self._safe_block_image_source(block) - if source: - return self._render_image(source, alt="table", class_name="mineru-table-image") - return _render_raw_fallback(content) - return f'
{html.escape(normalized_content, quote=False)}
' - source = self._safe_block_image_source(block) - return self._render_image(source, alt="table", class_name="mineru-table-image") if source else "" - - def _render_chart_block(self, block: ChartBlock) -> str: - """按原始子块顺序渲染 chart 图片、结构内容及说明。""" - parts: list[str] = [] - for child in block.content: - if isinstance(child, ChartBodyBlock): - body = self._render_chart_body(block, child) - elif isinstance(child, ChartAnnotationBlock): - body = self._render_annotation(child) - else: - raise TypeError(f"Unsupported chart child: {type(child).__name__}") - if body: - parts.append(body) - return f'
{"".join(parts)}
' if parts else "" - - def _render_chart_body(self, parent: ChartBlock, block: ChartBodyBlock) -> str: - """渲染 chart 图片，并将并存结构内容放入 details。""" - source = self._safe_block_image_source(block) - content = self._render_chart_content(block.content) - parts = [self._render_image(source, alt=parent.sub_type or "chart", class_name="mineru-chart-image")] if source else [] - if content.html: - if source: - parts.append(self._render_details(content.html, parent.sub_type or "chart content")) - else: - parts.append(content.html) - return "".join(parts) - - def _render_chart_content(self, content: str) -> HtmlInlineResult: - """按 HTML、严格 GFM 表格、普通行内内容的顺序渲染 chart content。""" - normalized = content.strip() - if not normalized: - return HtmlInlineResult("") - if is_supported_html_markup(normalized): - return self._render_embedded_content(normalized) - if looks_like_gfm_table(normalized): - rendered = render_gfm_table_html(normalized) - if rendered is None: - return HtmlInlineResult(_render_raw_fallback(content)) - self._observe_inline(rendered) - return rendered - rendered = render_inline_content_html(content) - self._observe_inline(rendered) - return rendered - - def _render_code_block(self, block: CodeBlock) -> str: - """按原始子块顺序渲染代码或算法及其说明。""" - parts: list[str] = [] - for child in block.content: - if isinstance(child, CodeBodyBlock): - body = self._render_code_body(block, child) - elif isinstance(child, CodeAnnotationBlock): - body = self._render_annotation(child) - else: - raise TypeError(f"Unsupported code child: {type(child).__name__}") - if body: - parts.append(body) - return f'
{"".join(parts)}
' if parts else "" - - def _render_code_body(self, parent: CodeBlock, block: CodeBodyBlock) -> str: - """代码使用 Prism class，算法使用共享 inline AST 保留空白。""" - if parent.sub_type == BlockType.CODE: - escaped = html.escape(_replace_html_controls(block.content), quote=False) - if not block.content: - return '
' - language = _normalize_prism_language(parent.guess_lang) - class_attr = f' class="language-{language}"' if language else "" - pre_class = f"mineru-code language-{language}" if language else "mineru-code" - self.has_prism = bool(language) or self.has_prism - return f'
{escaped}
' - if parent.sub_type == RAW_ALGORITHM: - rendered = render_inline_nodes_html( - parse_inline_content(block.content), - linkify_text=False, - separate_adjacent_math=True, - preserve_newlines=True, - ) - self._observe_inline(rendered) - return f'
{rendered.html}
' if rendered.html else "" - raise ValueError(f"Unsupported code subtype: {parent.sub_type}") - - def _render_annotation( - self, - block: ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock, - ) -> str: - """把视觉说明渲染为保持原类型的独立段落。""" - rendered = render_inline_content_html(block.content) - self._observe_inline(rendered) - if not rendered.html: - return "" - is_caption = str(block.type).endswith("caption") - role_class = "mineru-caption" if is_caption else "mineru-footnote" - type_class = html.escape(str(block.type).replace("_", "-"), quote=True) - return f'

{rendered.html}

' - - def _render_embedded_content(self, content: str, *, linkify_text: bool = True) -> HtmlInlineResult: - """安全处理 body 富 HTML；普通内容走共享 inline AST。""" - normalized = content.strip() - if not normalized: - return HtmlInlineResult("") - if not is_supported_html_markup(normalized): - rendered = render_inline_nodes_html(parse_inline_content(content), linkify_text=linkify_text) - self._observe_inline(rendered) - return rendered - sanitized = sanitize_html_fragment(normalized, asset_base_url=self.asset_base_url) - rendered = self._replace_safe_equations(sanitized) - self._observe_inline(rendered) - return rendered - - def _replace_safe_equations(self, markup: str) -> HtmlInlineResult: - """把 sanitizer 保留的 eq 标签替换为可信 MathJax carrier。""" - if not markup: - return HtmlInlineResult("") - soup = BeautifulSoup(markup, "html.parser") - has_math = False - for equation in list(soup.find_all("eq")): - rendered = render_math_html(equation.get_text(), display=False) - if not rendered.html: - equation.decompose() - continue - replacement = BeautifulSoup(rendered.html, "html.parser") - equation.replace_with(*list(replacement.contents)) - has_math = True - return HtmlInlineResult(str(soup), has_math) - - def _safe_block_image_source(self, block: EquationBlock | ImageBodyBlock | TableBodyBlock | ChartBodyBlock) -> str | None: - """解析 block 图片来源，并应用 HTML renderer 的 URL 安全策略。""" - if block.image_path: - return sanitize_image_source(block.image_path, asset_base_url=self.asset_base_url) - if block.image_base64: - return sanitize_image_source(block.image_base64) - return None - - def _render_image(self, source: str, *, alt: str, class_name: str) -> str: - """构造属性已转义的安全 img 元素。""" - return ( - f'' - ) - - def _render_details(self, content: str, summary: str) -> str: - """构造浏览器原生折叠详情，不引入额外脚本。""" - return ( - '
' - f"{html.escape(_replace_html_controls(summary), quote=False)}{content}
" - ) - - def _observe_inline(self, rendered: HtmlInlineResult) -> None: - """累计当前文档是否需要加载 MathJax。""" - self.has_math = rendered.has_math or self.has_math - - def _render_standalone(self, article: str) -> str: - """用固定 CSS 和按需 CDN 资源包装完整 HTML 文档。""" - title = html.escape( - _replace_html_controls(_resolve_document_title(self.middle_json, self.document_title)), - quote=False, - ) - styles = _load_html_styles() - dependencies = [] - if self.has_math: - dependencies.append(_mathjax_head()) - if self.has_prism: - dependencies.append(_prism_head()) - if self.has_mermaid: - dependencies.append(_mermaid_head()) - dependency_html = "\n".join(dependencies) - return ( - "\n" - '\n' - "\n" - '\n' - '\n' - f"{title}\n" - f"\n" - f"{dependency_html}\n" - "\n" - '\n' - f"{article}\n" - "\n" - "" - ) +from ..types import MiddleJson +from .contracts import RenderMode def render_html( @@ -638,304 +15,16 @@ def render_html( standalone: bool = True, document_title: str | None = None, ) -> str: - """把严格 MiddleJson 无副作用地渲染为 HTML 文档或单根 fragment。""" - if not isinstance(middle_json, MiddleJson): - raise TypeError("render_html expects a MiddleJson instance") - if not isinstance(mode, RenderMode): - raise TypeError("mode must be a RenderMode value") - if not isinstance(asset_base_url, str): - raise TypeError("asset_base_url must be a string") - if not isinstance(standalone, bool): - raise TypeError("standalone must be a bool") - if document_title is not None and not isinstance(document_title, str): - raise TypeError("document_title must be a string or None") - return _HtmlRenderer( + """惰性加载 HTML 实现并渲染严格 MiddleJson。""" + from ._internal.html.renderer import render_html as _render_html + + return _render_html( middle_json, mode=mode, asset_base_url=asset_base_url, standalone=standalone, document_title=document_title, - ).render() - - -def _classify_list(items: list[ListItem], add_reference_bullets: bool) -> tuple[str, str | None, str]: - """根据直属 marker 选择原生列表类型或显式 marker 模式。""" - if add_reference_bullets: - return "ul", None, "mineru-list--reference" - if items and all(item.kind == "unordered" for item in items): - return "ul", None, "mineru-list--unordered" - if items and all(item.kind == "ordered" for item in items): - styles = {item.ordered_style for item in items} - if len(styles) == 1: - style = next(iter(styles)) - list_type = { - "lower-alpha": "a", - "upper-alpha": "A", - "lower-roman": "i", - "upper-roman": "I", - }.get(style or "") - return "ol", list_type, "mineru-list--ordered" - if items and all(item.kind == "none" for item in items): - return "ul", None, "mineru-list--unmarked" - return "ul", None, "mineru-list--explicit" - - -def _list_item_content( - item: ListItem, - add_reference_bullets: bool, - *, - explicit_markers: bool, -) -> tuple[str, str | None]: - """决定一个列表项应剥离、保留还是显式显示源 marker。""" - if add_reference_bullets: - if item.kind == "unordered": - return item.body, None - original = f"{item.leading}{item.marker or ''}{item.separator}{item.body}" - return original, None - if explicit_markers: - return item.body, item.marker - if item.kind in {"unordered", "ordered"}: - return item.body, None - return item.body, item.marker - - -def _serialize_index_items(items: list[dict[str, object]]) -> str: - """把目录项及其已挂接的嵌套 ul 序列化为 li。""" - rendered: list[str] = [] - for item in items: - nested = item["nested"] - assert isinstance(nested, list) - content = str(item["content"]) - if not content and not nested: - continue - class_attr = ' class="mineru-list-item--orphan"' if item["orphan"] else "" - rendered.append(f"{content}{''.join(str(value) for value in nested)}") - return "".join(rendered) - - -def _collect_title_anchor_ids(middle_json: MiddleJson) -> dict[str, str]: - """为正文首次出现的非空 anchor 分配无空白且 document-wide 唯一的 HTML id。""" - anchor_ids: dict[str, str] = {} - used_ids: set[str] = set() - for page in middle_json.pages: - for block in page.blocks: - if not isinstance(block, TitleBlockBase): - continue - anchor = _anchor_key(block.anchor) - visible_title = inline_plain_text(parse_inline_content(block.content)).strip() - if not anchor or not visible_title or anchor in anchor_ids: - continue - base_id = _html_anchor_base(anchor) - candidate = base_id - suffix = 2 - while candidate in used_ids: - candidate = f"{base_id}-{suffix}" - suffix += 1 - anchor_ids[anchor] = candidate - used_ids.add(candidate) - return anchor_ids - - -def _plain_content_text(content: str) -> str: - """从 body 内容提取安全 img alt 所需的可见纯文本。""" - if not content: - return "" - if is_supported_html_markup(content): - return BeautifulSoup(content, "html.parser").get_text(" ", strip=True) - return inline_plain_text(parse_inline_content(content)).strip() - - -def _extract_mermaid_flowchart_source(content: str) -> str | None: - """提取受限 Mermaid flowchart fence，拒绝其他图类型和可改写配置的语法。""" - normalized = _replace_html_controls(content).replace("\r\n", "\n").replace("\r", "\n").strip() - if not normalized: - return None - lines = normalized.split("\n") - if len(lines) < 3: - return None - opening = re.fullmatch(r"(?P\x60{3,})[ \t]*mermaid[ \t]*", lines[0], re.IGNORECASE) - if opening is None: - return None - fence_length = len(opening.group("fence")) - if re.fullmatch(rf"\x60{{{fence_length},}}[ \t]*", lines[-1]) is None: - return None - source = "\n".join(lines[1:-1]).strip("\n") - if not source or len(source) > _MERMAID_MAX_TEXT_SIZE: - return None - if source.lstrip().startswith("---") or "%%{" in source: - return None - first_statement = next( - (line.strip() for line in source.splitlines() if line.strip() and not line.lstrip().startswith("%%")), - "", ) - if _MERMAID_FLOWCHART_HEADER_RE.match(first_statement) is None: - return None - return source - - -def _contains_usable_table(markup: str) -> bool: - """确认清洗后的 markup 至少保留一个含单元格的 table。""" - soup = BeautifulSoup(markup, "html.parser") - return any(table.find(("td", "th")) is not None for table in soup.find_all("table")) - - -def _render_raw_fallback(content: str) -> str: - """把无法安全结构化的原内容转义为可见 pre，避免静默丢失。""" - return f'
{html.escape(_replace_html_controls(content), quote=False)}
' - - -def _normalize_prism_language(language: str | None) -> str | None: - """把模型语言名归一化为不可构造路径穿越的 Prism component 名。""" - normalized = (language or "").strip().lower() - normalized = _PRISM_LANGUAGE_ALIASES.get(normalized, normalized.replace("_", "-")) - if not normalized or normalized in _BLOCKED_PRISM_LANGUAGE_NAMES: - return None - return normalized if _SAFE_PRISM_LANGUAGE_RE.fullmatch(normalized) else None - - -def _resolve_document_title(middle_json: MiddleJson, explicit_title: str | None) -> str: - """按显式参数、首个正文文档标题、固定回退值解析 HTML title。""" - if explicit_title and explicit_title.strip(): - return explicit_title.strip() - for page in middle_json.pages: - for block in page.blocks: - if isinstance(block, DocTitleBlock): - title = inline_plain_text(parse_inline_content(block.content)).strip() - if title: - return title - return "MinerU Document" - - -def _anchor_key(anchor: str | None) -> str: - """保留 producer anchor 身份，仅去除首尾空白供标题与目录精确匹配。""" - return (anchor or "").strip() - - -def _html_anchor_base(anchor: str) -> str: - """把原始 anchor 转成无空白的 HTML id 基值，碰撞由 document registry 处理。""" - return re.sub(r"\s+", "-", _replace_html_controls(anchor)) - - -def _replace_html_controls(content: str) -> str: - """替换 HTML 不允许的 C0 控制字符与不可编码的孤立 surrogate。""" - return re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f\ud800-\udfff]", "\ufffd", content) - - -@lru_cache(maxsize=1) -def _load_html_styles() -> str: - """从包资源读取并缓存压缩后的 MinerU 独立样式。""" - root = resources.files("mineru").joinpath("resources", "html") - return root.joinpath(_STYLE_RESOURCE_NAME).read_text(encoding="utf-8") - - -def _mathjax_head() -> str: - """返回固定、收紧公式输入能力的 MathJax 4.1.2 head 片段。""" - return f""" -""" - - -def _prism_head() -> str: - """返回固定 Prism core、Autoloader 和当前文档局部高亮脚本。""" - return f""" - -""" - - -def _mermaid_head() -> str: - """返回固定 Mermaid 入口和逐图安全渲染、失败回退脚本。""" - return f""" -""" __all__ = ["render_html"] diff --git a/mineru/render/markdown.py b/mineru/render/markdown.py index 073c9a26..72ef201d 100644 --- a/mineru/render/markdown.py +++ b/mineru/render/markdown.py @@ -1,16 +1,11 @@ # Copyright (c) Opendatalab. All rights reserved. -"""严格 MiddleJson 到 Markdown 的公共渲染实现。""" +"""严格 MiddleJson 到 Markdown 的轻量公共门面。""" from __future__ import annotations -from ..config import LatexDelimitersConfig, config -from ..types import PAGE_AUXILIARY_BLOCK_TYPES, MiddleJson -from ._internal.common.planner import PlannedBlock, build_render_plan -from ._internal.markdown.blocks import render_planned_block +from ..types import MiddleJson from .contracts import ImageRenderer, RenderMode -_PAGE_SEPARATOR = "\n\n---\n\n" - def render_markdown( middle_json: MiddleJson, @@ -19,53 +14,15 @@ def render_markdown( asset_base_url: str = "", image_renderer: ImageRenderer | None = None, ) -> str: - """把严格 MiddleJson 纯函数式渲染为 Markdown 字符串。""" - if not isinstance(middle_json, MiddleJson): - raise TypeError("render_markdown expects a MiddleJson instance") - if not isinstance(mode, RenderMode): - raise TypeError("mode must be a RenderMode value") + """惰性加载 Markdown 实现并渲染严格 MiddleJson。""" + from ._internal.markdown.renderer import render_markdown as _render_markdown - delimiters = config.render.latex_delimiters - planned_pages = build_render_plan(middle_json, mode) - rendered_pages = [ - _render_page( - page, - mode=mode, - delimiters=delimiters, - asset_base_url=asset_base_url, - image_renderer=image_renderer, - ) - for page in planned_pages - ] - if mode is RenderMode.FULL: - return _PAGE_SEPARATOR.join(rendered_pages) - return "\n\n".join(page for page in rendered_pages if page) - - -def _render_page( - planned_blocks: list[PlannedBlock], - *, - mode: RenderMode, - delimiters: LatexDelimitersConfig, - asset_base_url: str, - image_renderer: ImageRenderer | None = None, -) -> str: - """渲染单页逻辑块，并在默认模式中过滤重复页元素。""" - rendered: list[str] = [] - for planned in planned_blocks: - if planned.removed: - continue - if mode is RenderMode.DEFAULT and planned.block.type in PAGE_AUXILIARY_BLOCK_TYPES: - continue - text = render_planned_block( - planned, - delimiters=delimiters, - asset_base_url=asset_base_url, - image_renderer=image_renderer, - ) - if text and text.strip(): - rendered.append(text.strip("\n")) - return "\n\n".join(rendered) + return _render_markdown( + middle_json, + mode=mode, + asset_base_url=asset_base_url, + image_renderer=image_renderer, + ) __all__ = ["render_markdown"] diff --git a/mineru/render/structured_content.py b/mineru/render/structured_content.py index da19b358..18473d9b 100644 --- a/mineru/render/structured_content.py +++ b/mineru/render/structured_content.py @@ -1,56 +1,11 @@ # Copyright (c) Opendatalab. All rights reserved. -"""严格 MiddleJson 到树形 Markdown structured_content 的公共渲染实现。""" +"""严格 MiddleJson 到 Structured Content 的轻量公共门面。""" from __future__ import annotations -from typing import Any, TypeAlias +from typing import Any -from mineru.config import LatexDelimitersConfig, config -from mineru.render._internal.markdown.blocks import ( - render_single_block, - render_title_inline_content, - render_visual_annotation, - render_visual_body_content, -) -from mineru.render._internal.markdown.assets import normalize_image_source, resolve_image_source -from mineru.render._internal.markdown.escaping import escape_standalone_marker_rule, escape_text_block_markdown_prefix -from mineru.types import ( - BlockType, - ChartAnnotationBlock, - ChartBlock, - ChartBodyBlock, - CodeAnnotationBlock, - CodeBlock, - DocTitleBlock, - EquationBlock, - ImageAnnotationBlock, - ImageBlock, - ImageBodyBlock, - ImagePayloadBlock, - MiddleJson, - PageBlock, - ParagraphTitleBlock, - TableAnnotationBlock, - TableBlock, - TableBodyBlock, -) - -VisualBlock: TypeAlias = ImageBlock | TableBlock | ChartBlock | CodeBlock -VisualAnnotationBlock: TypeAlias = ImageAnnotationBlock | TableAnnotationBlock | ChartAnnotationBlock | CodeAnnotationBlock - -_CAPTION_TYPES = { - BlockType.IMAGE_CAPTION, - BlockType.TABLE_CAPTION, - BlockType.CHART_CAPTION, - BlockType.CODE_CAPTION, -} -_FOOTNOTE_TYPES = { - BlockType.IMAGE_FOOTNOTE, - BlockType.TABLE_FOOTNOTE, - BlockType.CHART_FOOTNOTE, - BlockType.CODE_FOOTNOTE, -} -_REMOVED_BLOCK_FIELDS = {"content", "index", "guess_lang", "image_path", "image_base64"} +from ..types import MiddleJson def render_structured_content( @@ -58,124 +13,10 @@ def render_structured_content( *, asset_base_url: str = "", ) -> dict[str, Any]: - """把严格 MiddleJson 无副作用地渲染为树形 Markdown structured_content。""" - if not isinstance(middle_json, MiddleJson): - raise TypeError("render_structured_content expects a MiddleJson instance") + """惰性加载 Structured Content 实现并渲染严格 MiddleJson。""" + from ._internal.structured_content.renderer import render_structured_content as _render_structured_content - delimiters = config.render.latex_delimiters - document_fields = middle_json.model_dump( - mode="json", - exclude={"pages"}, - exclude_defaults=True, - ) - pages = [ - { - "page_idx": page.page_idx, - "blocks": [ - _render_content_block( - block, - delimiters=delimiters, - asset_base_url=asset_base_url, - ) - for block in page.blocks - ], - } - for page in middle_json.pages - ] - return {"pages": pages, **document_fields} - - -def _render_content_block( - block: PageBlock, - *, - delimiters: LatexDelimitersConfig, - asset_base_url: str, -) -> dict[str, Any]: - """保留父块元数据，并把 block 内容收敛为 Markdown 字符串。""" - payload = block.model_dump( - mode="json", - exclude=_REMOVED_BLOCK_FIELDS, - exclude_defaults=True, - ) - if isinstance(block, (DocTitleBlock, ParagraphTitleBlock)): - content = render_title_inline_content(block, delimiters) - payload["content"] = escape_standalone_marker_rule(escape_text_block_markdown_prefix(content)) - return payload - if isinstance(block, EquationBlock): - payload["content"] = block.content.strip() - image_source = _resolve_content_image_source(block, asset_base_url) - if image_source is not None: - payload["image_source"] = image_source - return payload - if not isinstance(block, (ImageBlock, TableBlock, ChartBlock, CodeBlock)): - payload["content"] = render_single_block( - block, - delimiters=delimiters, - asset_base_url=asset_base_url, - ) - return payload - - payload["content"] = render_visual_body_content( - block, - delimiters=delimiters, - asset_base_url=asset_base_url, - ) - payload["captions"] = _render_annotation_group(block, _CAPTION_TYPES, delimiters) - payload["footnotes"] = _render_annotation_group(block, _FOOTNOTE_TYPES, delimiters) - image_source = _resolve_visual_image_source(block, asset_base_url) - if image_source is not None: - payload["image_source"] = image_source - return payload - - -def _render_annotation_group( - block: VisualBlock, - accepted_types: set[str], - delimiters: LatexDelimitersConfig, -) -> list[dict[str, Any]]: - """按源 index 稳定排序视觉说明，并保留可用 bbox 与 Markdown 内容。""" - annotations: list[tuple[int, VisualAnnotationBlock]] = [] - for position, child in enumerate(block.content): - if ( - isinstance( - child, - (ImageAnnotationBlock, TableAnnotationBlock, ChartAnnotationBlock, CodeAnnotationBlock), - ) - and child.type in accepted_types - ): - annotations.append((position, child)) - annotations.sort(key=_annotation_sort_key) - rendered_annotations: list[dict[str, Any]] = [] - for _, child in annotations: - annotation_payload: dict[str, Any] = {} - if child.bbox is not None: - annotation_payload["bbox"] = list(child.bbox) - annotation_payload["content"] = render_visual_annotation(child, delimiters) - rendered_annotations.append(annotation_payload) - return rendered_annotations - - -def _annotation_sort_key( - item: tuple[int, VisualAnnotationBlock], -) -> tuple[bool, int, int]: - """让有 index 的说明升序优先，缺失 index 的说明稳定排在末尾。""" - position, block = item - return block.index is None, block.index if block.index is not None else 0, position - - -def _resolve_visual_image_source(block: VisualBlock, asset_base_url: str) -> str | None: - """解析视觉 body 实际选择的图片来源，并返回安全的 Markdown 地址。""" - for child in block.content: - if not isinstance(child, (ImageBodyBlock, TableBodyBlock, ChartBodyBlock)): - continue - return _resolve_content_image_source(child, asset_base_url) - return None - - -def _resolve_content_image_source(block: ImagePayloadBlock, asset_base_url: str) -> str | None: - """把图片载荷收敛为 structured_content 中唯一且安全的 image_source。""" - source = resolve_image_source(block, asset_base_url) - return normalize_image_source(source) if source else None + return _render_structured_content(middle_json, asset_base_url=asset_base_url) __all__ = ["render_structured_content"] diff --git a/mineru/render/writer.py b/mineru/render/writer.py deleted file mode 100644 index 01c24385..00000000 --- a/mineru/render/writer.py +++ /dev/null @@ -1,63 +0,0 @@ -# Copyright (c) Opendatalab. All rights reserved. -import os -from abc import ABC, abstractmethod - - -class DataWriter(ABC): - @abstractmethod - def write(self, path: str, data: bytes) -> None: - """Write the data to the file. - - Args: - path (str): the target file where to write - data (bytes): the data want to write - """ - pass - - def write_string(self, path: str, data: str) -> None: - """Write the data to file, the data will be encoded to bytes. - - Args: - path (str): the target file where to write - data (str): the data want to write - """ - - def safe_encode(data: str, method: str): - try: - bit_data = data.encode(encoding=method, errors="replace") - return bit_data, True - except: # noqa - return None, False - - for method in ["utf-8", "ascii"]: - bit_data, flag = safe_encode(data, method) - if flag: - self.write(path, bit_data) - break - - -class FileBasedDataWriter(DataWriter): - def __init__(self, parent_dir: str = "") -> None: - """Initialized with parent_dir. - - Args: - parent_dir (str, optional): the parent directory that may be used within methods. Defaults to ''. - """ - self._parent_dir = parent_dir - - def write(self, path: str, data: bytes) -> None: - """Write file with data. - - Args: - path (str): the path of file, if the path is relative path, it will be joined with parent_dir. - data (bytes): the data want to write - """ - fn_path = path - if not os.path.isabs(fn_path) and len(self._parent_dir) > 0: - fn_path = os.path.join(self._parent_dir, path) - - if not os.path.exists(os.path.dirname(fn_path)) and os.path.dirname(fn_path) != "": - os.makedirs(os.path.dirname(fn_path), exist_ok=True) - - with open(fn_path, "wb") as f: - f.write(data)