【中文】PDF文档切分\切片\拆分最优方案-数据预处理阶段，为后续导入RAG向量数据库和ES数据库实现双路召回

目的

将PDF文档拆开，拆开后每个数据是文档中的某一段，目的是保证每条数据都有较完整的语义，并且长度不会太长

项目自述

看了很多切分项目，包括langchain、Langchain-Chatchat、、Chinese-LangChain、LangChain-ChatGLM-Webui、ChatPDF、semchunk等等，效果还行，但是不够完美，毕竟他们的对"\n"的优先级设置的较高，使用pymupdf得到的文本中充斥着大量的"\n"，如果全部删掉也十分影响语义

切分逻辑

1、保持段落完整性

2、保持语义完整性

代码逻辑

1、转换PDF文件为DOCX文件

2、循环遍历paragraphs保持段落完整性

3、以句号为节点，保持语义完整性

代码实现

python 复制代码

import re
import os
import csv
from pdf2docx import Converter
from docx import Document

_NON_WHITESPACE_SEMANTIC_SPLITTERS = (
    '。\n\n','。\n','。', '？', '！'
    '；','\n','\r'  '）', '"', ''', '】', '......', # Sentence terminators.
    '?', '!', '*', # Sentence terminators.
    ';', '，', '(', ')', '[', ']', """, """, ''', ''', "'", '"', '`', # Clause separators.
    ':', '---', '...', # Sentence interrupters.
    '/', '\\', '--', '&', '-', # Word joiners.
)

def split_text(text, chunk_size=400):
    # 按照标点符号优先级逐级尝试进行切分
    for splitter in _NON_WHITESPACE_SEMANTIC_SPLITTERS:
        if splitter:
            parts = text.split(splitter)
            chunks = []
            current_chunk = ''
            for part in parts:
                if len(current_chunk) + len(part) + len(splitter) <= chunk_size:
                    current_chunk += part + splitter
                else:
                    if current_chunk:
                        chunks.append(current_chunk.strip())
                    current_chunk = part + splitter
            if current_chunk:
                chunks.append(current_chunk.strip())
            # 如果切分后的块总大小符合预期，则返回结果
            if all(len(chunk) <= chunk_size for chunk in chunks):
                return chunks

    # 如果所有标点符号都无法进行有效切分，则按固定大小切分
    return [text[i:i + chunk_size] for i in range(0, len(text), chunk_size)]



def pdf_to_docx(pdf_file_path):
    try:

        docx_path = os.path.join(os.path.dirname(pdf_file_path), os.path.basename(pdf_file_path).split(".")[0] +".docx")
        cv = Converter(pdf_file_path)
        cv.convert(docx_path)
        cv.close()
        return docx_path
    except Exception as e:
        print(f"转换过程中发生错误：{str(e)}")
        return False
    
def pdf2docx_to_csv(pdf_file_path, max_length=400):
    
    docx_path = pdf_to_docx(pdf_file_path)
    if not docx_path:
    	return False
    docx = Document(docx_path)

    result = []
    current_text = ""

    for paragraph in docx.paragraphs:
        sections = paragraph.text.strip()
        sections = re.sub(r'\s+', ' ', sections)
        sections = re.sub(r'(.)\1{4,}', r'\1', sections)
        if len(sections) > max_length:
            chunk_size = int(len(sections) / ((len(sections) // max_length) + 1))
            sections = split_text(sections, chunk_size=chunk_size)
        if isinstance(sections,str):
            sections = [sections,]
        for section in sections:
            if not current_text or len(current_text) + len(section) + 1 <= max_length:
                current_text += " " + section
            else:
                period_index = current_text.rfind('。')
                if period_index != -1:
                    period_text = current_text[:period_index+1].strip()
                    result.append(period_text)
                    current_text = current_text[period_index+1:].strip() + section
                else:
                    result.append(current_text.strip())
                    current_text = section
    if current_text.strip():
        result.append(current_text.strip())
    output_path = os.path.join(os.path.dirname(pdf_file_path), os.path.basename(pdf_file_path).split(".")[0] + "_pdf2docx_"+ ".csv")

    with open(output_path, 'w', newline='', encoding='utf-8') as csvfile:
        csvwriter = csv.writer(csvfile)
        csvwriter.writerow(['filename', 'text'])
        csvwriter.writerows(result)

    print(f"{pdf_file_path} 处理完成")


if __name__ == "__main__":
    pdf_file_path = "/path/to/your/xxx.pdf"
    pdf2docx_to_csv(pdf_file_path)

如果觉得好用就点个赞！