PolyglotPDF

Demo

🎬 Watch Full Video

llms has been added as the translation api of choice, Doubao ,Qwen ,deepseek v3 , gpt4-o-mini are recommended. The color space error can be resolved by filling the white areas in PDF files. The old text to text translation api has been removed.

In addition, consider adding arxiv search function and rendering arxiv papers after latex translation.

Pasges show

Chinese LLM API Application

Doubao & Deepseek

Apply through Volcengine platform:

• Application URL: Volcengine-Doubao
• Available Models: Doubao, Deepseek series models

Tongyi Qwen

Apply through Alibaba Cloud platform:

• Application URL: Alibaba Cloud-Tongyi Qwen
• Available Models: Qwen-Max, Qwen-Plus series models

Overview

PolyglotPDF is an advanced PDF processing tool that employs specialized techniques for ultra-fast text, table, and formula recognition in PDF documents, typically completing processing within 1 second. It features OCR capabilities and layout-preserving translation, with full document translations usually completed within 10 seconds (speed may vary depending on the translation API provider).

Features

• Ultra-Fast Recognition: Processes text, tables, and formulas in PDFs within ~1 second
• Layout-Preserving Translation: Maintains original document formatting while translating content
• OCR Support: Handles scanned documents efficiently
• Text-based PDF：No GPU required
• Quick Translation: Complete PDF translation in approximately 10 seconds
• Flexible API Integration: Compatible with various translation service providers
• Web-based Comparison Interface: Side-by-side comparison of original and translated documents
• Enhanced OCR Capabilities: Improved accuracy in text recognition and processing
• Support for offline translation: Use smaller translation model

Installation and Setup

1. Clone the repository:

git clone https://github.com/CBIhalsen/PolyglotPDF.git
cd polyglotpdf

2. Install required packages:

pip install -r requirements.txt

3. Configure your API key in config.json. The alicloud translation API is not recommended.

4. Run the application:

python app.py

5. Access the web interface: Open your browser and navigate to http://127.0.0.1:8000

Requirements

• Python 3.8+
• deepl==1.17.0
• Flask==2.0.1
• Flask-Cors==5.0.0
• langdetect==1.0.9
• Pillow==10.2.0
• PyMuPDF==1.24.0
• pytesseract==0.3.10
• requests==2.31.0
• tiktoken==0.6.0
• Werkzeug==2.0.1

Acknowledgments

This project leverages PyMuPDF's capabilities for efficient PDF processing and layout preservation.

Upcoming Improvements

• PDF chat functionality
• Academic PDF search integration
• Optimization for even faster processing speeds

Known Issues

• Issue Description: Error during text re-editing: code=4: only Gray, RGB, and CMYK colorspaces supported
• Symptom: Unsupported color space encountered during text block editing
• Current Workaround: Skip text blocks with unsupported color spaces
• Proposed Solution: Switch to OCR mode for entire pages containing unsupported color spaces
• Example: View PDF sample with unsupported color spaces

Font Optimization

Current font configuration in the start function of main.py:

# Current configuration
css=f"* {{font-family:{get_font_by_language(self.target_language)};font-size:auto;color: #111111 ;font-weight:normal;}}"

You can optimize font display through the following methods:

1. Modify Default Font Configuration

# Custom font styles
css=f"""* {{
    font-family: {get_font_by_language(self.target_language)};
    font-size: auto;
    color: #111111;
    font-weight: normal;
    letter-spacing: 0.5px;  # Adjust letter spacing
    line-height: 1.5;      # Adjust line height
}}"""

2. Embed Custom Fonts You can embed custom fonts by following these steps:

• Place font files (.ttf, .otf) in the project's fonts directory
• Use @font-face to declare custom fonts in CSS

css=f"""
@font-face {{
    font-family: 'CustomFont';
    src: url('fonts/your-font.ttf') format('truetype');
}}
* {{
    font-family: 'CustomFont', {get_font_by_language(self.target_language)};
    font-size: auto;
    font-weight: normal;
}}
"""

Basic Principles

This project follows similar basic principles as Adobe Acrobat DC's PDF editing, using PyMuPDF for text block recognition and manipulation:

• Core Process:

# Get text blocks from the page
blocks = page.get_text("dict")["blocks"]

# Process each text block
for block in blocks:
    if block.get("type") == 0:  # text block
        bbox = block["bbox"]     # get text block boundary
        text = ""
        font_info = None
        # Collect text and font information
        for line in block["lines"]:
            for span in line["spans"]:
                text += span["text"] + " "

This approach directly processes PDF text blocks, maintaining the original layout while achieving efficient text extraction and modification.

• Technical Choices:

  • Utilizes PyMuPDF for PDF parsing and editing
  • Focuses on text processing
  • Avoids complex operations like AI formula recognition, table processing, or page restructuring

• Why Avoid Complex Processing:

  • AI recognition of formulas, tables, and PDF restructuring faces severe performance bottlenecks
  • Complex AI processing leads to high computational costs
  • Significantly increased processing time (potentially tens of seconds or more)
  • Difficult to deploy at scale with low costs in production environments
  • Not suitable for online services requiring quick response times

• Project Scope:

  • This project only serves to demonstrate the correct approach for layout-preserved PDF translation and AI-assisted PDF reading. Converting PDF files to markdown format for large language models to read, in my opinion, is not a wise approach.
  • Aims for optimal performance-to-cost ratio

• Performance:

  • PolyglotPDF API response time: ~1 second per page
  • Low computational resource requirements, suitable for scale deployment
  • High cost-effectiveness for commercial applications