SKILL.md
readonly只读
name
doc-parser
description
>
version
1.0
文档解析技能
概述
本技能利用 docling(IBM 最先进的文档理解库)实现高级文档解析。可解析复杂的 PDF、Word 文档和图像,同时保留结构、提取表格、图形,并处理多栏布局。
使用方法
- 提供要解析的文档
- 指定要提取的内容(文本、表格、图形等)
- 我将解析文档并返回结构化数据
示例提示:
- "解析此 PDF 并提取所有表格"
- "将这篇学术论文转换为结构化 Markdown"
- "提取此文档中的图形和标题"
- "解析此报告并保留文档结构"
领域知识
docling 基础
from docling.document_converter import DocumentConverter
# 初始化转换器
converter = DocumentConverter()
# 转换文档
result = converter.convert("document.pdf")
# 访问解析后的内容
doc = result.document
print(doc.export_to_markdown())
支持的格式
| 格式 | 扩展名 | 说明 |
|---|---|---|
| 原生和扫描件 | ||
| Word | .docx | 完整结构保留 |
| PowerPoint | .pptx | 幻灯片作为章节 |
| 图片 | .png, .jpg | OCR + 布局分析 |
| HTML | .html | 结构保留 |
基本用法
from docling.document_converter import DocumentConverter
# 创建转换器
converter = DocumentConverter()
# 转换单个文档
result = converter.convert("report.pdf")
# 访问文档
doc = result.document
# 导出选项
markdown = doc.export_to_markdown()
text = doc.export_to_text()
json_doc = doc.export_to_dict()
高级配置
from docling.document_converter import DocumentConverter
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions
# 配置管道
pipeline_options = PdfPipelineOptions()
pipeline_options.do_ocr = True
pipeline_options.do_table_structure = True
pipeline_options.table_structure_options.do_cell_matching = True
# 使用选项创建转换器
converter = DocumentConverter(
allowed_formats=[InputFormat.PDF, InputFormat.DOCX],
pdf_backend_options=pipeline_options
)
result = converter.convert("document.pdf")
文档结构
# 文档层次结构
doc = result.document
# 访问元数据
print(doc.name)
print(doc.origin)
# 遍历内容
for element in doc.iterate_items():
print(f"类型: {element.type}")
print(f"文本: {element.text}")
if element.type == "table":
print(f"行数: {len(element.data.table_cells)}")
提取表格
from docling.document_converter import DocumentConverter
import pandas as pd
def extract_tables(doc_path):
"""从文档中提取所有表格。"""
converter = DocumentConverter()
result = converter.convert(doc_path)
doc = result.document
tables = []
for element in doc.iterate_items():
if element.type == "table":
# 获取表格数据
table_data = element.export_to_dataframe()
tables.append({
'page': element.prov[0].page_no if element.prov else None,
'dataframe': table_data
})
return tables
# 使用示例
tables = extract_tables("report.pdf")
for i, table in enumerate(tables):
print(f"第 {i+1} 个表格,位于第 {table['page']} 页:")
print(table['dataframe'])
提取图形
def extract_figures(doc_path, output_dir):
"""提取图形及其标题。"""
import os
converter = DocumentConverter()
result = converter.convert(doc_path)
doc = result.document
figures = []
os.makedirs(output_dir, exist_ok=True)
for element in doc.iterate_items():
if element.type == "picture":
figure_info = {
'caption': element.caption if hasattr(element, 'caption') else None,
'page': element.prov[0].page_no if element.prov else None,
}
# 如果可用,保存图像
if hasattr(element, 'image'):
img_path = os.path.join(output_dir, f"figure_{len(figures)+1}.png")
element.image.save(img_path)
figure_info['path'] = img_path
figures.append(figure_info)
return figures
处理多栏布局
from docling.document_converter import DocumentConverter
def parse_multicolumn(doc_path):
"""解析多栏布局的文档。"""
converter = DocumentConverter()
result = converter.convert(doc_path)
doc = result.document
# docling 自动处理栏检测
# 文本按阅读顺序返回
structured_content = []
for element in doc.iterate_items():
content_item = {
'type': element.type,
'text': element.text if hasattr(element, 'text') else None,
'level': element.level if hasattr(element, 'level') else None,
}
# 如果可用,添加边界框
if element.prov:
content_item['bbox'] = element.prov[0].bbox
content_item['page'] = element.prov[0].page_no
structured_content.append(content_item)
return structured_content
导出格式
from docling.document_converter import DocumentConverter
converter = DocumentConverter()
result = converter.convert("document.pdf")
doc = result.document
# Markdown 导出
markdown = doc.export_to_markdown()
with open("output.md", "w") as f:
f.write(markdown)
# 纯文本
text = doc.export_to_text()
# JSON/dict 格式
json_doc = doc.export_to_dict()
# HTML 格式(如果支持)
# html = doc.export_to_html()
批量处理
from docling.document_converter import DocumentConverter
from pathlib import Path
from concurrent.futures import ThreadPoolExecutor
def batch_parse(input_dir, output_dir, max_workers=4):
"""并行解析多个文档。"""
input_path = Path(input_dir)
output_path = Path(output_dir)
output_path.mkdir(exist_ok=True)
converter = DocumentConverter()
def process_single(doc_path):
try:
result = converter.convert(str(doc_path))
md = result.document.export_to_markdown()
out_file = output_path / f"{doc_path.stem}.md"
with open(out_file, 'w') as f:
f.write(md)
return {'file': str(doc_path), 'status': 'success'}
except Exception as e:
return {'file': str(doc_path), 'status': 'error', 'error': str(e)}
docs = list(input_path.glob('*.pdf')) + list(input_path.glob('*.docx'))
with ThreadPoolExecutor(max_workers=max_workers) as executor:
results = list(executor.map(process_single, docs))
return results
最佳实践
- 使用合适的管道:根据文档类型进行配置
- 处理大型文档:必要时分块处理
- 验证表格提取:复杂表格可能需要复核
- 检查 OCR 质量:对扫描文档启用 OCR
- 缓存结果:存储解析后的文档以便复用
常见模式
学术论文解析器
def parse_academic_paper(pdf_path):
"""解析学术论文结构。"""
converter = DocumentConverter()
result = converter.convert(pdf_path)
doc = result.document
paper = {
'title': None,
'abstract': None,
'sections': [],
'references': [],
'tables': [],
'figures': []
}
current_section = None
for element in doc.iterate_items():
text = element.text if hasattr(element, 'text') else ''
if element.type == 'title':
paper['title'] = text
elif element.type == 'heading':
if 'abstract' in text.lower():
current_section = 'abstract'
elif 'reference' in text.lower():
current_section = 'references'
else:
paper['sections'].append({
'title': text,
'content': ''
})
current_section = 'section'
elif element.type == 'paragraph':
if current_section == 'abstract':
paper['abstract'] = text
elif current_section == 'section' and paper['sections']:
paper['sections'][-1]['content'] += text + '\n'
elif element.type == 'table':
paper['tables'].append({
'caption': element.caption if hasattr(element, 'caption') else None,
'data': element.export_to_dataframe() if hasattr(element, 'export_to_dataframe') else None
})
return paper
报告转结构化数据
def parse_business_report(doc_path):
"""将商业报告解析为结构化格式。"""
converter = DocumentConverter()
result = converter.convert(doc_path)
doc = result.document
report = {
'metadata': {
'title': None,
'date': None,
'author': None
},
'executive_summary': None,
'sections': [],
'key_metrics': [],
'recommendations': []
}
# 解析文档结构
for element in doc.iterate_items():
# 根据文档结构实现解析逻辑
pass
return report
示例
示例 1:解析财务报告
from docling.document_converter import DocumentConverter
def parse_financial_report(pdf_path):
"""从财务报告中提取结构化数据。"""
converter = DocumentConverter()
result = converter.convert(pdf_path)
doc = result.document
financial_data = {
'income_statement': None,
'balance_sheet': None,
'cash_flow': None,
'notes': []
}
# 提取表格
tables = []
for element in doc.iterate_items():
if element.type == 'table':
table_df = element.export_to_dataframe()
# 识别表格类型
if 'revenue' in str(table_df).lower() or 'income' in str(table_df).lower():
financial_data['income_statement'] = table_df
elif 'asset' in str(table_df).lower() or 'liabilities' in str(table_df).lower():
financial_data['balance_sheet'] = table_df
elif 'cash' in str(table_df).lower():
financial_data['cash_flow'] = table_df
else:
tables.append(table_df)
# 提取 Markdown 作为注释
financial_data['markdown'] = doc.export_to_markdown()
return financial_data
report = parse_financial_report('annual_report.pdf')
print("利润表:")
print(report['income_statement'])
示例 2:技术文档解析器
from docling.document_converter import DocumentConverter
def parse_technical_docs(doc_path):
"""解析技术文档。"""
converter = DocumentConverter()
result = converter.convert(doc_path)
doc = result.document
documentation = {
'title': None,
'version': None,
'sections': [],
'code_blocks': [],
'diagrams': []
}
current_section = None
for element in doc.iterate_items():
if element.type == 'title':
documentation['title'] = element.text
elif element.type == 'heading':
current_section = {
'title': element.text,
'level': element.level if hasattr(element, 'level') else 1,
'content': []
}
documentation['sections'].append(current_section)
elif element.type == 'code':
if current_section:
current_section['content'].append({
'type': 'code',
'content': element.text
})
documentation['code_blocks'].append(element.text)
elif element.type == 'picture':
documentation['diagrams'].append({
'page': element.prov[0].page_no if element.prov else None,
'caption': element.caption if hasattr(element, 'caption') else None
})
return documentation
docs = parse_technical_docs('api_documentation.pdf')
print(f"标题: {docs['title']}")
print(f"章节数: {len(docs['sections'])}")
示例 3:合同分析
from docling.document_converter import DocumentConverter
def analyze_contract(pdf_path):
"""解析合同文档以提取关键条款。"""
converter = DocumentConverter()
result = converter.convert(pdf_path)
doc = result.document
contract = {
'parties': [],
'clauses': [],
'dates': [],
'amounts': [],
'full_text': doc.export_to_text()
}
import re
# 提取日期
date_pattern = r'\b\d{1,2}[/-]\d{1,2}[/-]\d{2,4}\b|\b(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)[a-z]* \d{1,2},? \d{4}\b'
contract['dates'] = re.findall(date_pattern, contract['full_text'], re.IGNORECASE)
# 提取货币金额
amount_pattern = r'\$[\d,]+(?:\.\d{2})?|\b\d+(?:,\d{3})*(?:\.\d{2})?\s*(?:USD|dollars)\b'
contract['amounts'] = re.findall(amount_pattern, contract['full_text'], re.IGNORECASE)
# 将章节解析为条款
for element in doc.iterate_items():
if element.type == 'heading':
contract['clauses'].append({
'title': element.text,
'content': ''
})
elif element.type == 'paragraph' and contract['clauses']:
contract['clauses'][-1]['content'] += element.text + '\n'
return contract
contract_data = analyze_contract('agreement.pdf')
print(f"关键日期: {contract_data['dates']}")
print(f"金额: {contract_data['amounts']}")
限制
- 非常大的文档可能需要分块处理
- 手写内容需要 OCR 预处理
- 复杂的嵌套表格可能需要手动复核
- 某些 PDF 类型(如加密)不受支持
- 建议使用 GPU 以获得最佳性能
安装
pip install docling
# 完整功能
pip install docling[all]
# OCR 支持
pip install docling[ocr]






