pdf-ocr-skill

SKILL.md

PDF OCR Skill

中文版本

PDF OCR技能用于从影印版PDF文件和图片文件中提取文字内容。该技能支持两种OCR引擎：

RapidOCR（本地引擎）：无需API密钥，免费使用，识别速度快

硅基流动大模型（云端引擎）：使用AI大模型进行高精度OCR识别

功能特性

支持影印版PDF文件的文字提取

支持多种图片格式的文字识别（JPG、PNG、BMP、GIF、TIFF、WEBP）

四引擎支持：RapidOCR（本地）、RapidDoc（增强）、PaddleOCR（本地）和硅基流动API（云端）

支持中文和英文文字识别

保持文字的顺序和结构

自动将PDF页面转换为图片进行识别

智能引擎切换：当RapidOCR初始化失败时自动切换到硅基流动API

安装
依赖要求

pip install pymupdf pillow requests python-dotenv

可选依赖（推荐）

安装RapidOCR以获得本地识别能力：

pip install rapidocr_onnxruntime

环境变量配置

复制 .env.example 文件并重命名为 .env

根据需要配置以下选项：

# OCR引擎选择
# - "rapid": 使用RapidOCR本地引擎（默认，无需API密钥）
# - "rapidoc": 使用RapidDoc增强引擎（无需API密钥）
# - "paddle": 使用PaddleOCR本地引擎（无需API密钥）
# - "siliconflow": 使用硅基流动API引擎（需要API密钥）
OCR_ENGINE=rapid

# 如果使用硅基流动API引擎，需要配置以下选项：
SILICON_FLOW_API_KEY=your_api_key_here
SILICON_FLOW_OCR_MODEL=deepseek-ai/DeepSeek-OCR

快速开始

使用默认引擎（RapidOCR本地识别）

# 导入OCR处理器
from scripts.pdf_ocr_processor import PDFOCRProcessor

# 创建处理器实例（默认使用RapidOCR）
processor = PDFOCRProcessor()

# 执行PDF OCR识别
result = processor.ocr_pdf('path/to/your/scanned.pdf')

# 获取识别结果
print(f"识别完成，共 {result['page_count']} 页")
print(f"使用引擎: {result['engine']}")
print(result['text'])

使用硅基流动API引擎

# 导入OCR处理器
from scripts.pdf_ocr_processor import PDFOCRProcessor

# 创建处理器实例，指定使用硅基流动API
processor = PDFOCRProcessor(engine="siliconflow")

# 执行PDF OCR识别
result = processor.ocr_pdf('path/to/your/scanned.pdf')

# 获取识别结果
print(f"识别完成，共 {result['page_count']} 页")
print(result['text'])

识别图片文件

# 导入OCR处理器
from scripts.pdf_ocr_processor import PDFOCRProcessor

# 创建处理器实例
processor = PDFOCRProcessor()  # 或 PDFOCRProcessor(engine="siliconflow")

# 执行图片OCR识别
result = processor.ocr_image_file('path/to/your/image.jpg')

# 获取识别结果
print(f"识别结果: {result['text']}")

命令行使用

# 使用默认RapidOCR引擎
python pdf_ocr_processor.py your_document.pdf

# 使用硅基流动API引擎
python pdf_ocr_processor.py your_document.pdf siliconflow

# 使用RapidDoc增强引擎
python pdf_ocr_processor.py your_document.pdf rapidoc

# 使用PaddleOCR引擎
python pdf_ocr_processor.py your_document.pdf paddle

进阶使用示例

批量处理多个PDF文件

import os
from scripts.pdf_ocr_processor import PDFOCRProcessor

# 创建处理器实例
processor = PDFOCRProcessor()

# 批量处理目录中的所有PDF文件
pdf_dir = "path/to/pdf/files"
output_dir = "path/to/output"
os.makedirs(output_dir, exist_ok=True)

for pdf_file in os.listdir(pdf_dir):
    if pdf_file.endswith('.pdf'):
        pdf_path = os.path.join(pdf_dir, pdf_file)
        output_path = os.path.join(output_dir, f"{os.path.splitext(pdf_file)[0]}.txt")
        
        print(f"处理文件: {pdf_file}")
        try:
            result = processor.ocr_pdf(pdf_path)
            
            # 保存识别结果到文本文件
            with open(output_path, 'w', encoding='utf-8') as f:
                f.write(f"=== PDF OCR 识别结果 ===\n")
                f.write(f"文件名: {pdf_file}\n")
                f.write(f"页数: {result['page_count']}\n")
                f.write(f"使用引擎: {result['engine']}\n\n")
                f.write(result['text'])
            
            print(f"处理完成，结果已保存到: {output_path}")
        except Exception as e:
            print(f"处理失败: {e}")

混合使用两种引擎

from scripts.pdf_ocr_processor import PDFOCRProcessor

def process_with_best_engine(pdf_path):
    """尝试使用RapidOCR，如果效果不佳则使用硅基流动API"""
    # 首先使用RapidOCR本地引擎
    rapid_processor = PDFOCRProcessor(engine="rapid")
    rapid_result = rapid_processor.ocr_pdf(pdf_path)
    
    # 简单评估识别效果（例如：检查识别出的文本长度）
    text_length = len(rapid_result['text'])
    
    if text_length < 100:  # 如果识别出的文本太短，可能效果不佳
        print("RapidOCR识别效果可能不佳，尝试使用硅基流动API...")
        silicon_processor = PDFOCRProcessor(engine="siliconflow")
        silicon_result = silicon_processor.ocr_pdf(pdf_path)
        return silicon_result
    else:
        return rapid_result

# 使用示例
result = process_with_best_engine('path/to/your/document.pdf')
print(f"识别完成，使用引擎: {result['engine']}")
print(result['text'])

支持的文件格式

PDF文件: .pdf

图片文件: .jpg, .jpeg, .png, .bmp, .gif, .tiff, .webp

输出格式

{
    "text": "识别的完整文本内容",
    "page_count": 页数,  # 图片文件始终为1
    "engine": "rapid" | "rapidoc" | "paddle" | "siliconflow"  # 使用的OCR引擎
}

使用场景

处理扫描版合同、协议等文档

提取影印版书籍、报告中的文字

处理无法直接复制文字的PDF文件

批量处理扫描版PDF文档

识别截图、扫描件等图片中的文字

处理手写体或印刷体图片文字识别

注意事项

RapidOCR引擎：

完全免费，无需网络连接

首次使用会自动下载模型文件

识别速度取决于CPU性能

硅基流动API引擎：

需要有效的API密钥

可能会产生费用

识别速度取决于文件页数、图片大小和网络状况

RapidDoc引擎：

完全免费，无需网络连接

支持版面分析、表格识别、公式识别和阅读顺序恢复

提供更结构化的输出，包括Markdown格式

处理时间可能比RapidOCR长，因为需要进行额外的分析

PaddleOCR引擎：

完全免费，无需网络连接

使用PP-OCRv5模型，具有较高的识别准确率

首次使用会自动下载模型文件

支持多种语言和场景的文字识别

对于复杂的扫描版PDF或图片，识别准确率可能会有所不同

建议使用高清晰度的扫描版PDF或图片以获得更好的识别效果

触发使用不同引擎的提示词

在与 AI IDE 中的助手交互时，您可以使用以下提示词来指定使用不同的 OCR 引擎：

📍 触发 RapidOCR（本地引擎）的提示词

"使用本地 OCR 引擎处理这个 PDF"

"用 RapidOCR 识别这个文件"

"本地处理，不需要 API"

"快速识别这个文档"

"离线处理这个 PDF"

"不使用硅基流动 API，用本地引擎"

📍 触发硅基流动 API（云端引擎）的提示词

"使用硅基流动 API 处理这个 PDF"

"用大模型 OCR 识别这个文件"

"高精度识别这个文档"

"处理复杂的扫描件"

"用云端 OCR 引擎"

"使用 AI 大模型识别"

📍 触发 RapidDoc（增强引擎）的提示词

"使用 RapidDoc 处理这个 PDF"

"用增强 OCR 识别这个文件"

"处理带有版面分析的 PDF"

"提取带有表格识别的文本"

"用 RapidDoc 获得更好的格式"

"增强 OCR 与版面分析"

📍 触发 PaddleOCR（本地引擎）的提示词

"使用 PaddleOCR 处理这个 PDF"

"用 PaddleOCR 识别这个文件"

"使用 PP-OCRv5 模型识别"

"用 PaddleOCR 进行高精度识别"

"PaddleOCR 本地处理"

"使用 PaddleOCR 引擎提取文字"

📍 示例对话

示例 1：使用本地引擎

用户：帮我处理这个扫描版 PDF，用本地 OCR 引擎快速识别
助手：好的，我将使用 RapidOCR 本地引擎为您处理。请提供 PDF 文件路径。

示例 2：使用云端引擎

用户：这个 PDF 包含手写体，需要高精度识别，用硅基流动 API
助手：理解，我将使用硅基流动 API 大模型为您处理。请提供 PDF 文件路径和您的 API 密钥（如果尚未配置）。

示例 3：自动选择

用户：帮我识别这个 PDF，选择最合适的引擎
助手：我将默认使用 RapidOCR 本地引擎为您处理。如果识别效果不理想，我们可以尝试使用硅基流动 API。

🔧 技术实现

当 AI 助手接收到这些提示词时，会：

解析用户意图，确定要使用的引擎

调用 PDFOCRProcessor(engine="rapid")、PDFOCRProcessor(engine="rapidoc")、PDFOCRProcessor(engine="paddle") 或 PDFOCRProcessor(engine="siliconflow")

执行 OCR 识别并返回结果

🎯 最佳实践

明确指定引擎：如果您对引擎有特定要求，最好在提示词中明确说明

提供上下文：说明文档类型（如手写体、复杂格式等）有助于助手选择合适的引擎

测试不同引擎：对于重要文档，可以尝试两种引擎并比较结果

通过使用这些提示词，您可以在与 AI IDE 交互时灵活控制 OCR 引擎的选择，获得最佳的识别效果

故障排除

常见问题及解决方案

RapidOCR初始化失败

问题：ModuleNotFoundError: No module named 'rapidocr_onnxruntime'

解决方案：安装RapidOCR依赖：pip install rapidocr_onnxruntime

硅基流动API 401错误

问题：Unauthorized: 401 Client Error

解决方案：检查API密钥是否正确配置在.env文件中

PDF转图片失败

问题：ImportError: No module named 'fitz'

解决方案：安装PyMuPDF依赖：pip install pymupdf

识别结果为空

问题：识别结果文本长度为0

解决方案：

检查PDF是否为扫描版（非文本PDF）

尝试使用硅基流动API引擎

确保PDF或图片清晰可读

English Version

PDF OCR Skill

PDF OCR Skill is used to extract text content from scanned PDF files and image files. This skill supports two OCR engines:

RapidOCR (local engine): No API key required, free to use, fast recognition speed

SiliconFlow Large Model (cloud engine): Uses AI large model for high-precision OCR recognition

Features

Support text extraction from scanned PDF files

Support text recognition from multiple image formats (JPG, PNG, BMP, GIF, TIFF, WEBP)

Quadruple-engine support: RapidOCR (local), RapidDoc (enhanced), PaddleOCR (local), and SiliconFlow API (cloud)

Support Chinese and English text recognition

Maintain text order and structure

Automatically convert PDF pages to images for recognition

Intelligent engine switching: automatically switch to SiliconFlow API when RapidOCR initialization fails

Installation

Dependencies

pip install pymupdf pillow requests python-dotenv

Optional Dependencies (Recommended)

Install RapidOCR for local recognition capability:

pip install rapidocr_onnxruntime

Environment Configuration

Copy .env.example file and rename it to .env

Configure the following options as needed:

# OCR engine selection
# - "rapid": Use RapidOCR local engine (default, no API key required)
# - "rapidoc": Use RapidDoc enhanced engine (no API key required)
# - "paddle": Use PaddleOCR local engine (no API key required)
# - "siliconflow": Use SiliconFlow API engine (API key required)
OCR_ENGINE=rapid

# If using SiliconFlow API engine, configure the following options:
SILICON_FLOW_API_KEY=your_api_key_here
SILICON_FLOW_OCR_MODEL=deepseek-ai/DeepSeek-OCR

Quick Start

Using Default Engine (RapidOCR Local Recognition)

# Import OCR processor
from scripts.pdf_ocr_processor import PDFOCRProcessor

# Create processor instance (default uses RapidOCR)
processor = PDFOCRProcessor()

# Perform PDF OCR recognition
result = processor.ocr_pdf('path/to/your/scanned.pdf')

# Get recognition result
print(f"Recognition completed, total {result['page_count']} pages")
print(f"Engine used: {result['engine']}")
print(result['text'])

Using SiliconFlow API Engine

# Import OCR processor
from scripts.pdf_ocr_processor import PDFOCRProcessor

# Create processor instance, specify to use SiliconFlow API
processor = PDFOCRProcessor(engine="siliconflow")

# Perform PDF OCR recognition
result = processor.ocr_pdf('path/to/your/scanned.pdf')

# Get recognition result
print(f"Recognition completed, total {result['page_count']} pages")
print(result['text'])

Recognizing Image Files

# Import OCR processor
from scripts.pdf_ocr_processor import PDFOCRProcessor

# Create processor instance
processor = PDFOCRProcessor()  # or PDFOCRProcessor(engine="siliconflow")

# Perform image OCR recognition
result = processor.ocr_image_file('path/to/your/image.jpg')

# Get recognition result
print(f"Recognition result: {result['text']}")

Command Line Usage

# Use default RapidOCR engine
python pdf_ocr_processor.py your_document.pdf

# Use SiliconFlow API engine
python pdf_ocr_processor.py your_document.pdf siliconflow

# Use RapidDoc enhanced engine
python pdf_ocr_processor.py your_document.pdf rapidoc

# Use PaddleOCR engine
python pdf_ocr_processor.py your_document.pdf paddle

Advanced Usage Examples

Batch Processing Multiple PDF Files

import os
from scripts.pdf_ocr_processor import PDFOCRProcessor

# Create processor instance
processor = PDFOCRProcessor()

# Batch process all PDF files in directory
pdf_dir = "path/to/pdf/files"
output_dir = "path/to/output"
os.makedirs(output_dir, exist_ok=True)

for pdf_file in os.listdir(pdf_dir):
    if pdf_file.endswith('.pdf'):
        pdf_path = os.path.join(pdf_dir, pdf_file)
        output_path = os.path.join(output_dir, f"{os.path.splitext(pdf_file)[0]}.txt")
        
        print(f"Processing file: {pdf_file}")
        try:
            result = processor.ocr_pdf(pdf_path)
            
            # Save recognition result to text file
            with open(output_path, 'w', encoding='utf-8') as f:
                f.write(f"=== PDF OCR Recognition Result ===\n")
                f.write(f"File name: {pdf_file}\n")
                f.write(f"Pages: {result['page_count']}\n")
                f.write(f"Engine used: {result['engine']}\n\n")
                f.write(result['text'])
            
            print(f"Processing completed, result saved to: {output_path}")
        except Exception as e:
            print(f"Processing failed: {e}")

Using Both Engines

from scripts.pdf_ocr_processor import PDFOCRProcessor

def process_with_best_engine(pdf_path):
    """Try using RapidOCR, if not good enough then use SiliconFlow API"""
    # First use RapidOCR local engine
    rapid_processor = PDFOCRProcessor(engine="rapid")
    rapid_result = rapid_processor.ocr_pdf(pdf_path)
    
    # Simple evaluation of recognition effect (e.g., check recognized text length)
    text_length = len(rapid_result['text'])
    
    if text_length < 100:  # If recognized text is too short, may not be good enough
        print("RapidOCR recognition effect may not be good enough, trying SiliconFlow API...")
        silicon_processor = PDFOCRProcessor(engine="siliconflow")
        silicon_result = silicon_processor.ocr_pdf(pdf_path)
        return silicon_result
    else:
        return rapid_result

# Usage example
result = process_with_best_engine('path/to/your/document.pdf')
print(f"Recognition completed, engine used: {result['engine']}")
print(result['text'])

Supported File Formats

PDF files: .pdf

Image files: .jpg, .jpeg, .png, .bmp, .gif, .tiff, .webp

Output Format

{
    "text": "Recognized full text content",
    "page_count": number_of_pages,  # Always 1 for image files
    "engine": "rapid" | "rapidoc" | "paddle" | "siliconflow"  # OCR engine used
}

Use Cases

Processing scanned contracts, agreements and other documents

Extracting text from photocopied books and reports

Processing PDF files with non-copyable text

Batch processing scanned PDF documents

Recognizing text in screenshots and scanned images

Processing handwritten or printed text in images

Notes

RapidOCR Engine:

Completely free, no network connection required

Model files will be automatically downloaded on first use

Recognition speed depends on CPU performance

SiliconFlow API Engine:

Requires a valid API key

May incur costs

Recognition speed depends on number of pages, image size, and network conditions

RapidDoc Engine:

Completely free, no network connection required

Supports layout analysis, table recognition, formula recognition, and reading order recovery

Provides more structured output including markdown format

Processing time may be longer than RapidOCR due to additional analysis

PaddleOCR Engine:

Completely free, no network connection required

Uses PP-OCRv5 model with high recognition accuracy

Model files will be automatically downloaded on first use

Supports text recognition for multiple languages and scenarios

Recognition accuracy may vary for complex scanned PDFs or images

It is recommended to use high-resolution scanned PDFs or images for better recognition results

Prompt Words for Different Engines

When interacting with assistants in AI IDEs, you can use the following prompt words to specify different OCR engines:

📍 Prompt Words for RapidOCR (Local Engine)

"Use local OCR engine to process this PDF"

"Recognize this file with RapidOCR"

"Local processing, no API needed"

"Quickly recognize this document"

"Process this PDF offline"

"Don't use SiliconFlow API, use local engine"

📍 Prompt Words for SiliconFlow API (Cloud Engine)

"Use SiliconFlow API to process this PDF"

"Recognize this file with large model OCR"

"High-precision recognition for this document"

"Process complex scanned documents"

"Use cloud OCR engine"

"Use AI large model for recognition"

📍 Prompt Words for RapidDoc (Enhanced Engine)

"Use RapidDoc to process this PDF"

"Recognize this file with enhanced OCR"

"Process PDF with layout analysis"

"Extract text with table recognition"

"Use RapidDoc for better formatting"

"Enhanced OCR with layout analysis"

📍 Prompt Words for PaddleOCR (Local Engine)

"Use PaddleOCR to process this PDF"

"Recognize this file with PaddleOCR"

"Use PP-OCRv5 model for recognition"

"Use PaddleOCR for high-precision recognition"

"PaddleOCR local processing"

"Extract text using PaddleOCR engine"

📍 Example Conversations

Example 1: Using Local Engine

User: Help me process this scanned PDF, use local OCR engine for quick recognition
Assistant: Sure, I'll use the RapidOCR local engine for you. Please provide the PDF file path.

Example 2: Using Cloud Engine

User: This PDF contains handwritten text, need high-precision recognition, use SiliconFlow API
Assistant: Understood, I'll use the SiliconFlow API large model for you. Please provide the PDF file path and your API key (if not already configured).

Example 3: Automatic Selection

User: Help me recognize this PDF, choose the most suitable engine
Assistant: I'll default to using the RapidOCR local engine for you. If the recognition effect is not ideal, we can try using SiliconFlow API.

🔧 Technical Implementation

When the AI assistant receives these prompt words, it will:

Parse the user's intent to determine the engine to use

Call PDFOCRProcessor(engine="rapid"), PDFOCRProcessor(engine="rapidoc"), PDFOCRProcessor(engine="paddle"), or PDFOCRProcessor(engine="siliconflow")

Execute OCR recognition and return the result

🎯 Best Practices

Clearly specify the engine: If you have specific requirements for the engine, it's best to clearly state it in the prompt

Provide context: Explaining the document type (e.g., handwritten, complex format) helps the assistant choose the appropriate engine

Test different engines: For important documents, you can try both engines and compare the results

By using these prompt words, you can flexibly control the OCR engine selection when interacting with AI IDEs to get the best recognition results

Troubleshooting

Common Issues and Solutions

RapidOCR Initialization Failure

Issue: ModuleNotFoundError: No module named 'rapidocr_onnxruntime'

Solution: Install RapidOCR dependency: pip install rapidocr_onnxruntime

SiliconFlow API 401 Error

Issue: Unauthorized: 401 Client Error

Solution: Check if the API key is correctly configured in the .env file

PDF to Image Conversion Failure

Issue: ImportError: No module named 'fitz'

Solution: Install PyMuPDF dependency: pip install pymupdf

Empty Recognition Result

Issue: Recognition result text length is 0

Solution:

Check if the PDF is a scanned version (non-text PDF)

Try using SiliconFlow API engine

Ensure the PDF or image is clear and readable

License

MIT License - See LICENSE.txt