pdf（PDF 文档处理）

PDF Processing Skill

Overview

This skill enables comprehensive PDF operations through Python libraries and command-line tools. Use it for reading, creating, modifying, and analyzing PDF documents.

Quick Start

from pypdf import PdfReader

reader = PdfReader("document.pdf")
for page in reader.pages:
    text = page.extract_text()
    print(text)

When to Use

• Converting PDFs to Markdown - Use OpenAI Codex for intelligent conversion (RECOMMENDED FIRST STEP)
• Extracting text and metadata from PDF files
• Merging multiple PDFs into a single document
• Splitting large PDFs into individual pages
• Adding watermarks or annotations to PDFs
• Password-protecting or decrypting PDFs
• Extracting images from PDF documents
• OCR processing for scanned documents
• Creating new PDFs with reportlab
• Extracting tables from structured PDFs

PDF to Markdown Conversion (OpenAI Codex)

IMPORTANT: For all PDF documents, utilize OpenAI Codex to convert contents to.md file first, then use the markdown for further work.

Why Convert to Markdown First?

• Better structure preservation - Maintains headings, lists, tables
• Easier text processing - Standard markdown format
• Improved AI understanding - Codex understands document structure
• Format flexibility - Markdown can be converted to any format
• Version control friendly - Plain text, diff-friendly

OpenAI Codex Conversion

Prerequisites:

pip install openai pypdf
export OPENAI_API_KEY="your-api-key-here"

Basic Conversion:

import openai
from pypdf import PdfReader
from pathlib import Path

def pdf_to_markdown_codex(pdf_path, output_md_path=None, model="gpt-4.1"):
    """
    Convert PDF to markdown using OpenAI Codex.

    Args:
        pdf_path: Path to PDF file
        output_md_path: Optional path for output .md file (auto-generated if None)
        model: OpenAI model to use (gpt-4.1, gpt-4.1-mini, etc.)

    Returns:
        Path to generated markdown file
    """
    # Extract text from PDF
    reader = PdfReader(pdf_path)
    pdf_text = ""

    for page_num, page in enumerate(reader.pages, 1):
        text = page.extract_text()
        pdf_text += f"\n\n--- Page {page_num} ---\n\n{text}"

    # Generate markdown using OpenAI Codex
    client = openai.OpenAI()

    response = client.chat.completions.create(
        model=model,
        messages=[
            {
                "role": "system",
                "content": """You are an expert document converter. Convert the provided PDF text
                to well-structured markdown format. Preserve:
                - Document structure (headings, sections)
                - Lists and bullet points
                - Tables (convert to markdown tables)
                - Code blocks and technical content
                - Links and references

                Format the output as clean, readable markdown."""
            },
            {
                "role": "user",
                "content": f"Convert this PDF text to markdown:\n\n{pdf_text}"
            }
        ],
        temperature=0.3,  # Lower temperature for more consistent formatting
    )

    markdown_content = response.choices[0].message.content

    # Save to file
    if output_md_path is None:
        pdf_stem = Path(pdf_path).stem
        output_md_path = Path(pdf_path).parent / f"{pdf_stem}.md"

    # Ensure parent directory exists
    Path(output_md_path).parent.mkdir(parents=True, exist_ok=True)
    Path(output_md_path).write_text(markdown_content, encoding='utf-8')

    return output_md_path

# Usage
md_file = pdf_to_markdown_codex("document.pdf")
print(f"Markdown saved to: {md_file}")

Batch Conversion:

from pathlib import Path
import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

def batch_pdf_to_markdown(pdf_directory, output_directory=None, model="gpt-4.1"):
    """
    Convert all PDFs in a directory to markdown.

    Args:
        pdf_directory: Directory containing PDF files
        output_directory: Optional output directory (defaults to pdf_directory/markdown)
        model: OpenAI model to use
    """
    pdf_dir = Path(pdf_directory)

    if output_directory is None:
        output_dir = pdf_dir / "markdown"
    else:
        output_dir = Path(output_directory)

    output_dir.mkdir(parents=True, exist_ok=True)

    pdf_files = list(pdf_dir.glob("*.pdf"))
    total = len(pdf_files)

    logger.info(f"Found {total} PDF files to convert")

    for i, pdf_file in enumerate(pdf_files, 1):
        try:
            output_md = output_dir / f"{pdf_file.stem}.md"

            logger.info(f"[{i}/{total}] Converting {pdf_file.name}...")
            pdf_to_markdown_codex(pdf_file, output_md, model=model)
            logger.info(f"✓ Saved to {output_md.name}")

        except Exception as e:
            logger.error(f"✗ Failed to convert {pdf_file.name}: {e}")

    logger.info(f"\nConversion complete! Files in: {output_dir}")

# Usage
batch_pdf_to_markdown("/path/to/pdfs", model="gpt-4.1")

Chunked Conversion for Large PDFs:

def pdf_to_markdown_chunked(pdf_path, output_md_path=None,
                            chunk_pages=10, model="gpt-4.1"):
    """
    Convert large PDF by processing in chunks.

    Args:
        pdf_path: Path to PDF file
        output_md_path: Optional output path
        chunk_pages: Number of pages per chunk
        model: OpenAI model to use
    """
    reader = PdfReader(pdf_path)
    total_pages = len(reader.pages)

    markdown_sections = []

    for start_page in range(0, total_pages, chunk_pages):
        end_page = min(start_page + chunk_pages, total_pages)

        # Extract chunk
        chunk_text = ""
        for page_num in range(start_page, end_page):
            text = reader.pages[page_num].extract_text()
            chunk_text += f"\n\n--- Page {page_num + 1} ---\n\n{text}"

        # Convert chunk
        client = openai.OpenAI()
        response = client.chat.completions.create(
            model=model,
            messages=[
                {
                    "role": "system",
                    "content": "Convert PDF text to markdown. Maintain structure and formatting."
                },
                {
                    "role": "user",
                    "content": f"Convert pages {start_page + 1}-{end_page} to markdown:\n\n{chunk_text}"
                }
            ],
            temperature=0.3,
        )

        markdown_sections.append(response.choices[0].message.content)
        print(f"Processed pages {start_page + 1}-{end_page}/{total_pages}")

    # Combine sections
    full_markdown = "\n\n---\n\n".join(markdown_sections)

    # Save
    if output_md_path is None:
        output_md_path = Path(pdf_path).with_suffix('.md')

    # Ensure parent directory exists
    Path(output_md_path).parent.mkdir(parents=True, exist_ok=True)
    Path(output_md_path).write_text(full_markdown, encoding='utf-8')

    return output_md_path

# Usage
md_file = pdf_to_markdown_chunked("large_document.pdf", chunk_pages=20)

Workflow: PDF → Markdown → Further Processing:

from pathlib import Path

def pdf_workflow(pdf_path):
    """
    Complete workflow: PDF → Markdown → Process markdown.

    Returns:
        dict with paths to original PDF, markdown, and processed content
    """
    # Step 1: Convert PDF to markdown using Codex
    print("Step 1: Converting PDF to markdown...")
    md_path = pdf_to_markdown_codex(pdf_path)

    # Step 2: Read markdown for further processing
    print("Step 2: Reading markdown content...")
    markdown_content = Path(md_path).read_text(encoding='utf-8')

    # Step 3: Further processing (example: extract headings)
    print("Step 3: Processing markdown...")
    headings = [line for line in markdown_content.split('\n') if line.startswith('#')]

    # Step 4: Additional analysis
    word_count = len(markdown_content.split())

    return {
        'pdf_path': pdf_path,
        'markdown_path': md_path,
        'markdown_content': markdown_content,
        'headings': headings,
        'word_count': word_count,
    }

# Usage
result = pdf_workflow("technical_document.pdf")
print(f"Markdown saved: {result['markdown_path']}")
print(f"Found {len(result['headings'])} headings")
print(f"Word count: {result['word_count']}")

# Now work with the markdown
with open(result['markdown_path']) as f:
    markdown = f.read()
    # Do further processing with clean markdown

Cost-Effective Options:

# Use GPT-4.1-mini for cost savings
md_file = pdf_to_markdown_codex("document.pdf", model="gpt-4.1-mini")

# Or use local extraction + Codex for formatting only
from pypdf import PdfReader

def hybrid_conversion(pdf_path):
    """Extract text locally, use Codex only for formatting."""
    # Extract text (free)
    reader = PdfReader(pdf_path)
    raw_text = ""
    for page in reader.pages:
        raw_text += page.extract_text()

    # Use Codex just for markdown formatting (lower cost)
    client = openai.OpenAI()
    response = client.chat.completions.create(
        model="gpt-4.1-mini",
        messages=[
            {
                "role": "system",
                "content": "Format the text as markdown. Add appropriate headings, lists, and structure."
            },
            {
                "role": "user",
                "content": raw_text
            }
        ],
        temperature=0.3,
    )

    markdown = response.choices[0].message.content
    output_path = Path(pdf_path).with_suffix('.md')
    output_path.write_text(markdown, encoding='utf-8')

    return output_path

Best Practices:

1. Always convert to markdown first - Makes downstream processing easier
2. Use chunking for large PDFs - Avoids token limits and API timeouts
3. Cache conversions - Store markdown files to avoid re-conversion
4. Choose model based on complexity - GPT-4.1 for complex docs, GPT-4.1-mini for simple ones
5. Validate output - Check that markdown structure makes sense
6. Handle errors gracefully - Log failures, continue batch processing

CLI Tool:

#!/usr/bin/env python3
"""PDF to Markdown converter using OpenAI Codex."""

import argparse
import logging
from pathlib import Path
import openai
from pypdf import PdfReader

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

def pdf_to_markdown_codex(pdf_path, output_md_path=None, model="gpt-4.1"):
    """
    Convert PDF to markdown using OpenAI Codex.

    Args:
        pdf_path: Path to PDF file
        output_md_path: Optional path for output .md file (auto-generated if None)
        model: OpenAI model to use (gpt-4.1, gpt-4.1-mini, etc.)

    Returns:
        Path to generated markdown file
    """
    # Extract text from PDF
    reader = PdfReader(pdf_path)
    pdf_text = ""

    for page_num, page in enumerate(reader.pages, 1):
        text = page.extract_text()
        pdf_text += f"\n\n--- Page {page_num} ---\n\n{text}"

    # Generate markdown using OpenAI Codex
    client = openai.OpenAI()

    response = client.chat.completions.create(
        model=model,
        messages=[
            {
                "role": "system",
                "content": """You are an expert document converter. Convert the provided PDF text
                to well-structured markdown format. Preserve:
                - Document structure (headings, sections)
                - Lists and bullet points
                - Tables (convert to markdown tables)
                - Code blocks and technical content
                - Links and references

                Format the output as clean, readable markdown."""
            },
            {
                "role": "user",
                "content": f"Convert this PDF text to markdown:\n\n{pdf_text}"
            }
        ],
        temperature=0.3,
    )

    markdown_content = response.choices[0].message.content

    # Save to file
    if output_md_path is None:
        pdf_stem = Path(pdf_path).stem
        output_md_path = Path(pdf_path).parent / f"{pdf_stem}.md"

    # Ensure parent directory exists
    Path(output_md_path).parent.mkdir(parents=True, exist_ok=True)
    Path(output_md_path).write_text(markdown_content, encoding='utf-8')

    return output_md_path

def batch_pdf_to_markdown(pdf_directory, output_directory=None, model="gpt-4.1"):
    """
    Convert all PDFs in a directory to markdown.

    Args:
        pdf_directory: Directory containing PDF files
        output_directory: Optional output directory (defaults to pdf_directory/markdown)
        model: OpenAI model to use
    """
    pdf_dir = Path(pdf_directory)

    if output_directory is None:
        output_dir = pdf_dir / "markdown"
    else:
        output_dir = Path(output_directory)

    output_dir.mkdir(parents=True, exist_ok=True)

    pdf_files = list(pdf_dir.glob("*.pdf"))
    total = len(pdf_files)

    logger.info(f"Found {total} PDF files to convert")

    for i, pdf_file in enumerate(pdf_files, 1):
        try:
            output_md = output_dir / f"{pdf_file.stem}.md"

            logger.info(f"[{i}/{total}] Converting {pdf_file.name}...")
            pdf_to_markdown_codex(pdf_file, output_md, model=model)
            logger.info(f"✓ Saved to {output_md.name}")

        except Exception as e:
            logger.error(f"✗ Failed to convert {pdf_file.name}: {e}")

    logger.info(f"\nConversion complete! Files in: {output_dir}")

def main():
    parser = argparse.ArgumentParser(description='Convert PDF to Markdown using OpenAI')
    parser.add_argument('input', help='PDF file or directory')
    parser.add_argument('-o', '--output', help='Output directory or file')
    parser.add_argument('-m', '--model', default='gpt-4.1', help='OpenAI model (gpt-4.1, gpt-4.1-mini)')
    parser.add_argument('--chunk-pages', type=int, default=10, help='Pages per chunk (unused in basic mode)')

    args = parser.parse_args()

    input_path = Path(args.input)

    if input_path.is_file():
        # Single file
        output = args.output or input_path.with_suffix('.md')
        md_path = pdf_to_markdown_codex(input_path, output, model=args.model)
        print(f"✓ Converted: {md_path}")
    else:
        # Directory
        batch_pdf_to_markdown(input_path, args.output, model=args.model)

if __name__ == '__main__':
    main()

Save as pdf2md.py and use:

# Single file
python pdf2md.py document.pdf

# Directory
python pdf2md.py /path/to/pdfs -o /path/to/markdown

# With GPT-4.1-mini (cheaper)
python pdf2md.py document.pdf --model gpt-4.1-mini

Python Libraries

pypdf - Core PDF Operations

Merging PDFs:

from pypdf import PdfMerger

merger = PdfMerger()
merger.append("file1.pdf")
merger.append("file2.pdf")
merger.write("merged.pdf")
merger.close()

Splitting PDFs:

from pypdf import PdfReader, PdfWriter

reader = PdfReader("document.pdf")
for i, page in enumerate(reader.pages):
    writer = PdfWriter()
    writer.add_page(page)
    writer.write(f"page_{i+1}.pdf")

Extracting Metadata:

reader = PdfReader("document.pdf")
info = reader.metadata
print(f"Author: {info.author}")
print(f"Title: {info.title}")
print(f"Pages: {len(reader.pages)}")

pdfplumber - Advanced Text Extraction

Text with Layout Preservation:

import pdfplumber

with pdfplumber.open("document.pdf") as pdf:
    for page in pdf.pages:
        text = page.extract_text()
        print(text)

Table Extraction:

import pdfplumber

with pdfplumber.open("document.pdf") as pdf:
    page = pdf.pages[0]
    tables = page.extract_tables()
    for table in tables:
        for row in table:
            print(row)

reportlab - Creating PDFs

Create PDF from Scratch:

from reportlab.lib.pagesizes import letter
from reportlab.pdfgen import canvas

c = canvas.Canvas("output.pdf", pagesize=letter)
c.drawString(100, 750, "Hello, World!")
c.showPage()
c.save()

Multi-page Documents:

from reportlab.lib.pagesizes import letter
from reportlab.platypus import SimpleDocTemplate, Paragraph
from reportlab.lib.styles import getSampleStyleSheet

doc = SimpleDocTemplate("output.pdf", pagesize=letter)
styles = getSampleStyleSheet()
story = []

story.append(Paragraph("Title", styles['Heading1']))
story.append(Paragraph("Body text here.", styles['Normal']))

doc.build(story)

Command-Line Tools

pdftotext (Poppler)

pdftotext document.pdf output.txt
pdftotext -layout document.pdf output.txt  # Preserve layout

qpdf

# Merge PDFs
qpdf --empty --pages file1.pdf file2.pdf -- merged.pdf

# Split pages
qpdf document.pdf --pages . 1-5 -- first_five.pdf

# Decrypt
qpdf --decrypt encrypted.pdf decrypted.pdf

pdftk

# Merge
pdftk file1.pdf file2.pdf cat output merged.pdf

# Split
pdftk document.pdf burst output page_%02d.pdf

# Rotate
pdftk document.pdf cat 1-endeast output rotated.pdf

PDF-Large-Reader - Robust Extraction for Large Files

For large PDFs (100MB+, 1000+ pages), use the pdf-large-reader library with robust table extraction.

Why Use PDF-Large-Reader?

• Memory-efficient - Handles 100MB+ PDFs without memory issues
• Robust table extraction - Handles irregular tables with column count normalization
• Multiple output formats - Generator (streaming), List, or Plain Text
• Automatic strategy selection - Intelligent chunk size calculation
• Complete extraction - Text, images, tables, and metadata in one pass
• High test coverage - 93.58% coverage with 215 tests

Installation

# From the pdf-large-reader repository
cd /mnt/github/workspace-hub/pdf-large-reader
pip install -e .

# Or with extras
pip install -e ".[dev,progress]"

Quick Start

from pdf_large_reader import process_large_pdf, extract_text_only, extract_everything

# Simple text extraction
text = extract_text_only("large_document.pdf")
print(text)

# Process with automatic strategy selection
pages = process_large_pdf(
    "large_document.pdf",
    output_format="list",
    extract_images=True,
    extract_tables=True
)

# Memory-efficient streaming for very large files
for page in process_large_pdf("huge_file.pdf", output_format="generator"):
    print(f"Page {page.page_number}: {len(page.text)} characters")

Robust Table Extraction

NEW: Column Count Normalization (v1.3.0+)

The table extraction now handles irregular tables with different column counts:

from pdf_large_reader import extract_everything

# Extract everything including tables with robust error handling
pages = extract_everything("technical_standard.pdf")

for page in pages:
    if 'tables' in page.metadata:
        tables = page.metadata['tables']
        print(f"Page {page.page_number}: Found {len(tables)} tables")

        for i, table_df in enumerate(tables):
            print(f"  Table {i+1}: {table_df.shape[0]} rows x {table_df.shape[1]} cols")
            print(table_df.head())

How It Works:

• Detects table-like structures from text positioning
• Normalizes column counts across all rows
• Pads short rows with empty strings
• Gracefully handles malformed tables with try-except
• Logs warnings instead of crashing

Typical Performance:

• API Std 650 (28 MB, 461 pages): 14,648 chars/sec, 5.18 pages/sec
• API RP 579 (41 MB, 966 pages): 2,090 chars/sec, 8.48 pages/sec

Command Line Usage

# Extract text from PDF
pdf-large-reader document.pdf

# Save to file
pdf-large-reader document.pdf --output result.txt

# Extract with images and tables
pdf-large-reader document.pdf --extract-images --extract-tables

# Use generator format for large files
pdf-large-reader huge.pdf --output-format generator

# Verbose output
pdf-large-reader document.pdf --verbose

API Reference

# Main entry point with automatic strategy
process_large_pdf(
    pdf_path,
    output_format="generator",    # "generator" (default), "list", or "text"
    extract_images=False,         # Extract images
    extract_tables=False,         # Extract tables with normalization
    chunk_size=None,              # Auto-calculated if None
    fallback_api_key=None,        # OpenAI API key for complex pages
    fallback_model="gpt-4.1",      # Model for fallback extraction
    progress_callback=None,       # Progress tracking function
    auto_strategy=True            # Enable automatic strategy selection
)

# Quick text extraction
extract_text_only(pdf_path) -> str

# Extract with images
extract_pages_with_images(pdf_path) -> List[PDFPage]

# Extract with tables
extract_pages_with_tables(pdf_path) -> List[PDFPage]

# Extract everything
extract_everything(pdf_path) -> List[PDFPage]

PDFPage Data Structure

@dataclass
class PDFPage:
    page_number: int          # Page number (1-indexed)
    text: str                 # Extracted text from page
    images: List[dict]        # Extracted images with metadata
    metadata: dict            # Page metadata including tables

Performance Benchmarks

Tested on Ubuntu 22.04, Python 3.11, 16GB RAM:

Real-World Validation

Tested with actual API standards:

• ✅ API RP 579 (2000) - 41 MB, 966 pages
• ✅ API Std 650 (2001) - 28 MB, 461 pages
• ✅ All extraction methods working (text, auto strategy, generator, complete)
• ✅ Table extraction with column normalization
• ✅ Image extraction (461-966 images per document)

Common Tasks

OCR for Scanned Documents

import pytesseract
from pdf2image import convert_from_path

images = convert_from_path("scanned.pdf")
for i, image in enumerate(images):
    text = pytesseract.image_to_string(image)
    print(f"Page {i+1}:\n{text}")

Add Watermark

from pypdf import PdfReader, PdfWriter

reader = PdfReader("document.pdf")
watermark = PdfReader("watermark.pdf")
writer = PdfWriter()

for page in reader.pages:
    page.merge_page(watermark.pages[0])
    writer.add_page(page)

writer.write("watermarked.pdf")

Extract Images

from pypdf import PdfReader

reader = PdfReader("document.pdf")
for page_num, page in enumerate(reader.pages):
    for img_num, image in enumerate(page.images):
        with open(f"image_{page_num}_{img_num}.png", "wb") as f:
            f.write(image.data)

Password Protection

from pypdf import PdfReader, PdfWriter

reader = PdfReader("document.pdf")
writer = PdfWriter()

for page in reader.pages:
    writer.add_page(page)

writer.encrypt("user_password", "owner_password")
writer.write("protected.pdf")

Execution Checklist

• Verify input PDF exists and is readable
• Check if PDF is encrypted/DRM-protected
• Choose appropriate library for task (pypdf vs pdfplumber)
• Handle multi-page documents correctly
• Validate output file was created
• Clean up temporary files

Error Handling

Common Errors

Error: FileNotFoundError

• Cause: PDF file path is incorrect
• Solution: Verify file path and ensure file exists

Error: PdfReadError (encrypted)

• Cause: PDF is password-protected or DRM-encrypted
• Solution: Provide password or use qpdf to decrypt

Error: Empty text extraction

• Cause: PDF contains scanned images, not text
• Solution: Use OCR with pytesseract and pdf2image

Error: DependencyError (Tesseract)

• Cause: Tesseract OCR not installed
• Solution: sudo apt-get install tesseract-ocr or brew install tesseract

Metrics

Quick Reference

Dependencies

# Core PDF libraries
pip install pypdf pdfplumber reportlab pytesseract pdf2image

# OpenAI Codex for PDF to Markdown conversion
pip install openai

System tools:

• Poppler (pdftotext, pdftoppm)
• qpdf
• pdftk
• Tesseract OCR

Environment variables:

export OPENAI_API_KEY="your-api-key-here"

Version History

• 1.2.2 (2026-01-04): Fixed P2 issue - added parents=True to all mkdir() calls to handle nested output paths; prevents FileNotFoundError when creating directories with non-existent parent paths
• 1.2.1 (2026-01-04): Fixed CLI tool missing imports - added complete standalone script with all required imports (openai, pypdf, logging) and function definitions; resolved P1 issue from Codex review
• 1.2.0 (2026-01-04): MAJOR UPDATE - Added OpenAI Codex integration for PDF-to-Markdown conversion as recommended first step for all PDF processing; includes batch conversion, chunking for large files, cost-effective options, and complete CLI tool
• 1.1.0 (2026-01-02): Added Quick Start, When to Use, Execution Checklist, Error Handling, Metrics sections; updated frontmatter with version, category, related_skills
• 1.0.0 (2024-10-15): Initial release with pypdf, pdfplumber, reportlab, CLI tools