mineru 2.6.7__py3-none-any.whl → 2.7.0__py3-none-any.whl

mineru/backend/vlm/vlm_magic_model.py

@@ -6,7 +6,7 @@ from loguru import logger
 from mineru.utils.boxbase import calculate_overlap_area_in_bbox1_area_ratio
 from mineru.utils.enum_class import ContentType, BlockType
 from mineru.utils.guess_suffix_or_lang import guess_language_by_text
-from mineru.utils.magic_model_utils import reduct_overlap, tie_up_category_by_distance_v3
+from mineru.utils.magic_model_utils import reduct_overlap, tie_up_category_by_index
 
 
 class MagicModel:
@@ -317,13 +317,14 @@ def clean_content(content):
     return content
 
 
-def __tie_up_category_by_distance_v3(blocks, subject_block_type, object_block_type):
+def __tie_up_category_by_index(blocks, subject_block_type, object_block_type):
+    """基于index的主客体关联包装函数"""
     # 定义获取主体和客体对象的函数
     def get_subjects():
         return reduct_overlap(
             list(
                 map(
-                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"], "angle":x["angle"]},
+                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"], "angle": x["angle"]},
                     filter(
                         lambda x: x["type"] == subject_block_type,
                         blocks,
@@ -336,7 +337,7 @@ def __tie_up_category_by_distance_v3(blocks, subject_block_type, object_block_ty
         return reduct_overlap(
             list(
                 map(
-                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"], "angle":x["angle"]},
+                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"], "angle": x["angle"]},
                     filter(
                         lambda x: x["type"] == object_block_type,
                         blocks,
@@ -346,15 +347,15 @@ def __tie_up_category_by_distance_v3(blocks, subject_block_type, object_block_ty
         )
 
     # 调用通用方法
-    return tie_up_category_by_distance_v3(
+    return tie_up_category_by_index(
         get_subjects,
         get_objects
     )
 
 
 def get_type_blocks(blocks, block_type: Literal["image", "table", "code"]):
-    with_captions = __tie_up_category_by_distance_v3(blocks, f"{block_type}_body", f"{block_type}_caption")
-    with_footnotes = __tie_up_category_by_distance_v3(blocks, f"{block_type}_body", f"{block_type}_footnote")
+    with_captions = __tie_up_category_by_index(blocks, f"{block_type}_body", f"{block_type}_caption")
+    with_footnotes = __tie_up_category_by_index(blocks, f"{block_type}_body", f"{block_type}_footnote")
     ret = []
     for v in with_captions:
         record = {
@@ -368,49 +369,6 @@ def get_type_blocks(blocks, block_type: Literal["image", "table", "code"]):
     return ret
 
 
-def fix_two_layer_blocks_back(blocks, fix_type: Literal["image", "table", "code"]):
-    need_fix_blocks = get_type_blocks(blocks, fix_type)
-    fixed_blocks = []
-    not_include_blocks = []
-    processed_indices = set()
-
-    # 处理需要组织成two_layer结构的blocks
-    for block in need_fix_blocks:
-        body = block[f"{fix_type}_body"]
-        caption_list = block[f"{fix_type}_caption_list"]
-        footnote_list = block[f"{fix_type}_footnote_list"]
-
-        body["type"] = f"{fix_type}_body"
-        for caption in caption_list:
-            caption["type"] = f"{fix_type}_caption"
-            processed_indices.add(caption["index"])
-        for footnote in footnote_list:
-            footnote["type"] = f"{fix_type}_footnote"
-            processed_indices.add(footnote["index"])
-
-        processed_indices.add(body["index"])
-
-        two_layer_block = {
-            "type": fix_type,
-            "bbox": body["bbox"],
-            "blocks": [
-                body,
-            ],
-            "index": body["index"],
-        }
-        two_layer_block["blocks"].extend([*caption_list, *footnote_list])
-
-        fixed_blocks.append(two_layer_block)
-
-    # 添加未处理的blocks
-    for block in blocks:
-        if block["index"] not in processed_indices:
-            # 直接添加未处理的block
-            not_include_blocks.append(block)
-
-    return fixed_blocks, not_include_blocks
-
-
 def fix_two_layer_blocks(blocks, fix_type: Literal["image", "table", "code"]):
     need_fix_blocks = get_type_blocks(blocks, fix_type)
     fixed_blocks = []
@@ -418,25 +376,16 @@ def fix_two_layer_blocks(blocks, fix_type: Literal["image", "table", "code"]):
     processed_indices = set()
 
     # 特殊处理表格类型，确保标题在表格前，注脚在表格后
-    if fix_type == "table":
+    if fix_type in ["table", "image"]:
         # 收集所有不合适的caption和footnote
         misplaced_captions = []  # 存储(caption, 原始block索引)
         misplaced_footnotes = []  # 存储(footnote, 原始block索引)
 
-        # 第一步：移除不符合位置要求的caption和footnote
+        # 第一步：移除不符合位置要求的footnote
         for block_idx, block in enumerate(need_fix_blocks):
             body = block[f"{fix_type}_body"]
             body_index = body["index"]
 
-            # 检查caption应在body前或同位置
-            valid_captions = []
-            for caption in block[f"{fix_type}_caption_list"]:
-                if caption["index"] <= body_index:
-                    valid_captions.append(caption)
-                else:
-                    misplaced_captions.append((caption, block_idx))
-            block[f"{fix_type}_caption_list"] = valid_captions
-
             # 检查footnote应在body后或同位置
             valid_footnotes = []
             for footnote in block[f"{fix_type}_footnote_list"]:
@@ -446,28 +395,6 @@ def fix_two_layer_blocks(blocks, fix_type: Literal["image", "table", "code"]):
                     misplaced_footnotes.append((footnote, block_idx))
             block[f"{fix_type}_footnote_list"] = valid_footnotes
 
-        # 第二步：重新分配不合规的caption到合适的body
-        for caption, original_block_idx in misplaced_captions:
-            caption_index = caption["index"]
-            best_block_idx = None
-            min_distance = float('inf')
-
-            # 寻找索引大于等于caption_index的最近body
-            for idx, block in enumerate(need_fix_blocks):
-                body_index = block[f"{fix_type}_body"]["index"]
-                if body_index >= caption_index and idx != original_block_idx:
-                    distance = body_index - caption_index
-                    if distance < min_distance:
-                        min_distance = distance
-                        best_block_idx = idx
-
-            if best_block_idx is not None:
-                # 找到合适的body，添加到对应block的caption_list
-                need_fix_blocks[best_block_idx][f"{fix_type}_caption_list"].append(caption)
-            else:
-                # 没找到合适的body，作为普通block处理
-                not_include_blocks.append(caption)
-
         # 第三步：重新分配不合规的footnote到合适的body
         for footnote, original_block_idx in misplaced_footnotes:
             footnote_index = footnote["index"]
@@ -502,13 +429,22 @@ def fix_two_layer_blocks(blocks, fix_type: Literal["image", "table", "code"]):
                 caption_list.sort(key=lambda x: x["index"], reverse=True)
                 filtered_captions = [caption_list[0]]
                 for i in range(1, len(caption_list)):
-                    # 检查是否与前一个caption连续(降序所以是-1)
-                    if caption_list[i]["index"] == caption_list[i - 1]["index"] - 1:
+                    prev_index = caption_list[i - 1]["index"]
+                    curr_index = caption_list[i]["index"]
+
+                    # 检查是否连续
+                    if curr_index == prev_index - 1:
                         filtered_captions.append(caption_list[i])
                     else:
-                        # 出现gap,后续所有caption都作为普通block
-                        not_include_blocks.extend(caption_list[i:])
-                        break
+                        # 检查gap中是否只有body_index
+                        gap_indices = set(range(curr_index + 1, prev_index))
+                        if gap_indices == {body_index}:
+                            # gap中只有body_index,不算真正的gap
+                            filtered_captions.append(caption_list[i])
+                        else:
+                            # 出现真正的gap,后续所有caption都作为普通block
+                            not_include_blocks.extend(caption_list[i:])
+                            break
                 # 恢复升序
                 filtered_captions.reverse()
                 block[f"{fix_type}_caption_list"] = filtered_captions
@@ -592,7 +528,6 @@ def fix_list_blocks(list_blocks, text_blocks, ref_text_blocks):
     for list_block in list_blocks:
         # 统计list_block["blocks"]中所有block的type，用众数作为list_block的sub_type
         type_count = {}
-        line_content = []
         for sub_block in list_block["blocks"]:
             sub_block_type = sub_block["type"]
             if sub_block_type not in type_count:

mineru/backend/vlm/vlm_middle_json_mkcontent.py

@@ -1,8 +1,11 @@
 import os
 
 from loguru import logger
+
+from mineru.utils.char_utils import full_to_half_exclude_marks, is_hyphen_at_line_end
 from mineru.utils.config_reader import get_latex_delimiter_config, get_formula_enable, get_table_enable
 from mineru.utils.enum_class import MakeMode, BlockType, ContentType, ContentTypeV2
+from mineru.utils.language import detect_lang
 
 latex_delimiters_config = get_latex_delimiter_config()
 
@@ -18,9 +21,18 @@ display_right_delimiter = delimiters['display']['right']
 inline_left_delimiter = delimiters['inline']['left']
 inline_right_delimiter = delimiters['inline']['right']
 
+
 def merge_para_with_text(para_block, formula_enable=True, img_buket_path=''):
-    para_text = ''
+    block_text = ''
     for line in para_block['lines']:
+        for span in line['spans']:
+            if span['type'] in [ContentType.TEXT]:
+                span['content'] = full_to_half_exclude_marks(span['content'])
+                block_text += span['content']
+    block_lang = detect_lang(block_text)
+
+    para_text = ''
+    for i, line in enumerate(para_block['lines']):
         for j, span in enumerate(line['spans']):
             span_type = span['type']
             content = ''
@@ -34,17 +46,51 @@ def merge_para_with_text(para_block, formula_enable=True, img_buket_path=''):
                 else:
                     if span.get('image_path', ''):
                         content = f"![]({img_buket_path}/{span['image_path']})"
-            # content = content.strip()
+
+            content = content.strip()
             if content:
-                if span_type in [ContentType.TEXT, ContentType.INLINE_EQUATION]:
-                    if j == len(line['spans']) - 1:
+
+                if span_type == ContentType.INTERLINE_EQUATION:
+                    para_text += content
+                    continue
+
+                # 定义CJK语言集合(中日韩)
+                cjk_langs = {'zh', 'ja', 'ko'}
+                # logger.info(f'block_lang: {block_lang}, content: {content}')
+
+                # 判断是否为行末span
+                is_last_span = j == len(line['spans']) - 1
+
+                if block_lang in cjk_langs:  # 中文/日语/韩文语境下，换行不需要空格分隔,但是如果是行内公式结尾，还是要加空格
+                    if is_last_span and span_type != ContentType.INLINE_EQUATION:
                         para_text += content
                     else:
                         para_text += f'{content} '
-                elif span_type == ContentType.INTERLINE_EQUATION:
-                    para_text += content
+                else:
+                    # 西方文本语境下 每行的最后一个span判断是否要去除连字符
+                    if span_type in [ContentType.TEXT, ContentType.INLINE_EQUATION]:
+                        # 如果span是line的最后一个且末尾带有-连字符，那么末尾不应该加空格,同时应该把-删除
+                        if (
+                                is_last_span
+                                and span_type == ContentType.TEXT
+                                and is_hyphen_at_line_end(content)
+                        ):
+                            # 如果下一行的第一个span是小写字母开头，删除连字符
+                            if (
+                                    i+1 < len(para_block['lines'])
+                                    and para_block['lines'][i + 1].get('spans')
+                                    and para_block['lines'][i + 1]['spans'][0].get('type') == ContentType.TEXT
+                                    and para_block['lines'][i + 1]['spans'][0].get('content', '')
+                                    and para_block['lines'][i + 1]['spans'][0]['content'][0].islower()
+                            ):
+                                para_text += content[:-1]
+                            else:  # 如果没有下一行，或者下一行的第一个span不是小写字母开头，则保留连字符但不加空格
+                                para_text += content
+                        else:  # 西方文本语境下 content间需要空格分隔
+                            para_text += f'{content} '
     return para_text
 
+
 def mk_blocks_to_markdown(para_blocks, make_mode, formula_enable, table_enable, img_buket_path=''):
     page_markdown = []
     for para_block in para_blocks:
@@ -479,22 +525,76 @@ def get_body_data(para_block):
 
 
 def merge_para_with_text_v2(para_block):
-    para_content = []
-    para_type = para_block['type']
+    block_text = ''
     for line in para_block['lines']:
         for span in line['spans']:
+            if span['type'] in [ContentType.TEXT]:
+                span['content'] = full_to_half_exclude_marks(span['content'])
+                block_text += span['content']
+    block_lang = detect_lang(block_text)
+
+    para_content = []
+    para_type = para_block['type']
+    for i, line in enumerate(para_block['lines']):
+        for j, span in enumerate(line['spans']):
             span_type = span['type']
-            if span['content']:
-                if para_type == BlockType.PHONETIC and span_type == ContentTypeV2.SPAN_TEXT:
-                    span_type = ContentTypeV2.SPAN_PHONETIC
+            if span.get("content", '').strip():
+                if span_type == ContentType.TEXT:
+                    if para_type == BlockType.PHONETIC:
+                        span_type = ContentTypeV2.SPAN_PHONETIC
+                    else:
+                        span_type = ContentTypeV2.SPAN_TEXT
                 if span_type == ContentType.INLINE_EQUATION:
                     span_type = ContentTypeV2.SPAN_EQUATION_INLINE
                 if span_type in [
                     ContentTypeV2.SPAN_TEXT,
+                ]:
+                    # 定义CJK语言集合(中日韩)
+                    cjk_langs = {'zh', 'ja', 'ko'}
+                    # logger.info(f'block_lang: {block_lang}, content: {content}')
+
+                    # 判断是否为行末span
+                    is_last_span = j == len(line['spans']) - 1
+
+                    if block_lang in cjk_langs:  # 中文/日语/韩文语境下，换行不需要空格分隔,但是如果是行内公式结尾，还是要加空格
+                        if is_last_span:
+                            span_content = span['content']
+                        else:
+                            span_content = f"{span['content']} "
+                    else:
+                        # 如果span是line的最后一个且末尾带有-连字符，那么末尾不应该加空格,同时应该把-删除
+                        if (
+                                is_last_span
+                                and is_hyphen_at_line_end(span['content'])
+                        ):
+                            # 如果下一行的第一个span是小写字母开头，删除连字符
+                            if (
+                                    i + 1 < len(para_block['lines'])
+                                    and para_block['lines'][i + 1].get('spans')
+                                    and para_block['lines'][i + 1]['spans'][0].get('type') == ContentType.TEXT
+                                    and para_block['lines'][i + 1]['spans'][0].get('content', '')
+                                    and para_block['lines'][i + 1]['spans'][0]['content'][0].islower()
+                            ):
+                                span_content = span['content'][:-1]
+                            else:  # 如果没有下一行，或者下一行的第一个span不是小写字母开头，则保留连字符但不加空格
+                                span_content = span['content']
+                        else:
+                            # 西方文本语境下content间需要空格分隔
+                            span_content = f"{span['content']} "
+
+                    if para_content and para_content[-1]['type'] == span_type:
+                        # 合并相同类型的span
+                        para_content[-1]['content'] += span_content
+                    else:
+                        span_content = {
+                            'type': span_type,
+                            'content': span_content,
+                        }
+                        para_content.append(span_content)
+
+                elif span_type in [
                     ContentTypeV2.SPAN_PHONETIC,
                     ContentTypeV2.SPAN_EQUATION_INLINE,
-                    ContentTypeV2.SPAN_MD,
-                    ContentTypeV2.SPAN_CODE_INLINE,
                 ]:
                     span_content = {
                         'type': span_type,
@@ -534,13 +634,12 @@ def union_make(pdf_info_dict: list,
                 output_content.append(para_content)
         elif make_mode == MakeMode.CONTENT_LIST_V2:
             # https://github.com/drunkpig/llm-webkit-mirror/blob/dev6/docs/specification/output_format/content_list_spec.md
-            page_contents = []
             para_blocks = (paras_of_layout or []) + (paras_of_discarded or [])
-            if not para_blocks:
-                continue
-            for para_block in para_blocks:
-                para_content = make_blocks_to_content_list_v2(para_block, img_buket_path, page_size)
-                page_contents.append(para_content)
+            page_contents = []
+            if para_blocks:
+                for para_block in para_blocks:
+                    para_content = make_blocks_to_content_list_v2(para_block, img_buket_path, page_size)
+                    page_contents.append(para_content)
             output_content.append(page_contents)
 
     if make_mode in [MakeMode.MM_MD, MakeMode.NLP_MD]:

mineru/cli/client.py

@@ -1,10 +1,15 @@
 # Copyright (c) Opendatalab. All rights reserved.
 import os
+import sys
+
 import click
 from pathlib import Path
 from loguru import logger
 
-from mineru.utils.check_sys_env import is_mac_os_version_supported
+log_level = os.getenv("MINERU_LOG_LEVEL", "INFO").upper()
+logger.remove()  # 移除默认handler
+logger.add(sys.stderr, level=log_level)  # 添加新handler
+
 from mineru.utils.cli_parser import arg_parse
 from mineru.utils.config_reader import get_device
 from mineru.utils.guess_suffix_or_lang import guess_suffix_by_path
@@ -13,10 +18,6 @@ from ..version import __version__
 from .common import do_parse, read_fn, pdf_suffixes, image_suffixes
 
 
-backends = ['pipeline', 'vlm-transformers', 'vlm-vllm-engine', 'vlm-lmdeploy-engine', 'vlm-http-client']
-if is_mac_os_version_supported():
-    backends.append("vlm-mlx-engine")
-
 @click.command(context_settings=dict(ignore_unknown_options=True, allow_extra_args=True))
 @click.pass_context
 @click.version_option(__version__,
@@ -50,24 +51,23 @@ if is_mac_os_version_supported():
       txt: Use text extraction method.
       ocr: Use OCR method for image-based PDFs.
     Without method specified, 'auto' will be used by default.
-    Adapted only for the case where the backend is set to 'pipeline'.""",
+    Adapted only for the case where the backend is set to 'pipeline' and 'hybrid-*'.""",
     default='auto',
 )
 @click.option(
     '-b',
     '--backend',
     'backend',
-    type=click.Choice(backends),
+    type=click.Choice(['pipeline', 'vlm-http-client', 'hybrid-http-client', 'vlm-auto-engine', 'hybrid-auto-engine',]),
     help="""\b
     the backend for parsing pdf:
       pipeline: More general.
-      vlm-transformers: More general, but slower.
-      vlm-mlx-engine: Faster than transformers(macOS 13.5+).
-      vlm-vllm-engine: Faster(vllm-engine).
-      vlm-lmdeploy-engine: Faster(lmdeploy-engine).
-      vlm-http-client: Faster(client suitable for openai-compatible servers).
-    Without method specified, pipeline will be used by default.""",
-    default='pipeline',
+      vlm-auto-engine: High accuracy via local computing power.
+      vlm-http-client: High accuracy via remote computing power(client suitable for openai-compatible servers).
+      hybrid-auto-engine: Next-generation high accuracy solution via local computing power.
+      hybrid-http-client: High accuracy but requires a little local computing power(client suitable for openai-compatible servers).
+    Without method specified, hybrid-auto-engine will be used by default.""",
+    default='hybrid-auto-engine',
 )
 @click.option(
     '-l',
@@ -78,7 +78,7 @@ if is_mac_os_version_supported():
     help="""
     Input the languages in the pdf (if known) to improve OCR accuracy.
     Without languages specified, 'ch' will be used by default.
-    Adapted only for the case where the backend is set to "pipeline".
+    Adapted only for the case where the backend is set to 'pipeline' and 'hybrid-*'.
     """,
     default='ch',
 )
@@ -88,7 +88,7 @@ if is_mac_os_version_supported():
     'server_url',
     type=str,
     help="""
-    When the backend is `vlm-http-client`, you need to specify the server_url, for example:`http://127.0.0.1:30000`
+    When the backend is `<vlm/hybrid>-http-client`, you need to specify the server_url, for example:`http://127.0.0.1:30000`
     """,
     default=None,
 )
@@ -130,7 +130,7 @@ if is_mac_os_version_supported():
     'device_mode',
     type=str,
     help="""Device mode for model inference, e.g., "cpu", "cuda", "cuda:0", "npu", "npu:0", "mps".
-         Adapted only for the case where the backend is set to "pipeline" and "vlm-transformers". """,
+         Adapted only for the case where the backend is set to "pipeline". """,
     default=None,
 )
 @click.option(