doctra

Package Overview

Dependencies

Maintainers

Versions

Alerts

File Explorer

Advanced tools

License

Install Socket

Detect and block malicious and high-risk dependencies

Install

doctra

Parse, extract, and analyze documents with ease

PyPI

Version: 0.9.7

Maintainers: 1

🚀 Doctra - Document Parser Library 📑🔎

Doctra Logo

📋 Table of Contents

🛠️ Installation

From PyPI (recommended)

pip install doctra

From source

git clone https://github.com/AdemBoukhris457/Doctra.git
cd Doctra
pip install .

System Dependencies

Doctra requires Poppler for PDF processing. Install it based on your operating system:

Ubuntu/Debian

sudo apt install poppler-utils

macOS

brew install poppler

Windows

Download and install from Poppler for Windows or use conda:

conda install -c conda-forge poppler

Google Colab

!sudo apt install poppler-utils

⚡ Quick Start

from doctra.parsers.structured_pdf_parser import StructuredPDFParser

# Initialize the parser
parser = StructuredPDFParser()

# Parse a PDF document
parser.parse("path/to/your/document.pdf")

🔧 Core Components

StructuredPDFParser

The StructuredPDFParser is a comprehensive PDF parser that extracts all types of content from PDF documents. It processes PDFs through layout detection, extracts text using OCR, saves images for visual elements, and optionally converts charts/tables to structured data using Vision Language Models (VLM).

Key Features:

Layout Detection: Uses PaddleOCR for accurate document layout analysis
OCR Processing: Supports both PyTesseract (default) and PaddleOCR PP-OCRv5_server for text extraction
Visual Element Extraction: Saves figures, charts, and tables as images
VLM Integration: Optional conversion of visual elements to structured data
Multiple Output Formats: Generates Markdown, Excel, and structured JSON

Basic Usage:

from doctra.parsers.structured_pdf_parser import StructuredPDFParser

# Basic parser without VLM (uses default PyTesseract OCR engine)
parser = StructuredPDFParser()

# Parser with VLM for structured data extraction
from doctra.engines.vlm.service import VLMStructuredExtractor

# Initialize VLM engine
vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",  # or "gemini", "anthropic", "openrouter", "qianfan", "ollama"
    api_key="your_api_key_here"
)

# Pass VLM engine to parser
parser = StructuredPDFParser(vlm=vlm_engine)

# Parse document
parser.parse("document.pdf")

OCR Engine Configuration:

Doctra uses a dependency injection pattern for OCR engines. You initialize the OCR engine externally and pass it to the parser:

from doctra.parsers.structured_pdf_parser import StructuredPDFParser
from doctra.engines.ocr import PytesseractOCREngine, PaddleOCREngine

# Option 1: Use default PyTesseract (automatic if ocr_engine=None)
parser = StructuredPDFParser()  # Creates default PyTesseractOCREngine internally

# Option 2: Explicitly configure PyTesseract
tesseract_ocr = PytesseractOCREngine(
    lang="eng",      # Language code
    psm=4,           # Page segmentation mode
    oem=3,           # OCR engine mode
    extra_config=""  # Additional Tesseract config
)
parser = StructuredPDFParser(ocr_engine=tesseract_ocr)

# Option 3: Use PaddleOCR for better accuracy
paddle_ocr = PaddleOCREngine(
    device="gpu",                          # "gpu" or "cpu"
    use_doc_orientation_classify=False,    # Document orientation detection
    use_doc_unwarping=False,              # Text image rectification
    use_textline_orientation=False        # Text line orientation
)
parser = StructuredPDFParser(ocr_engine=paddle_ocr)

# Option 4: Reuse OCR engine across multiple parsers
shared_ocr = PytesseractOCREngine(lang="eng", psm=6, oem=3)
parser1 = StructuredPDFParser(ocr_engine=shared_ocr)
parser2 = EnhancedPDFParser(ocr_engine=shared_ocr)  # Reuse same instance

VLM Engine Configuration:

Doctra uses the same dependency injection pattern for VLM engines. You initialize the VLM engine externally and pass it to the parser:

from doctra.parsers.structured_pdf_parser import StructuredPDFParser
from doctra.engines.vlm.service import VLMStructuredExtractor

# Option 1: No VLM (default)
parser = StructuredPDFParser()  # VLM processing disabled

# Option 2: Initialize VLM engine and pass to parser
vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",  # or "gemini", "anthropic", "openrouter", "qianfan", "ollama"
    vlm_model="gpt-5",      # Optional, uses default if None
    api_key="your_api_key"
)
parser = StructuredPDFParser(vlm=vlm_engine)

# Option 3: Reuse VLM engine across multiple parsers
shared_vlm = VLMStructuredExtractor(
    vlm_provider="gemini",
    api_key="your_api_key"
)
parser1 = StructuredPDFParser(vlm=shared_vlm)
parser2 = EnhancedPDFParser(vlm=shared_vlm)  # Reuse same instance
parser3 = ChartTablePDFParser(vlm=shared_vlm)  # Reuse same instance

Advanced Configuration:

from doctra.engines.ocr import PytesseractOCREngine, PaddleOCREngine

# Option 1: Using PyTesseract (default)
ocr_engine = PytesseractOCREngine(
    lang="eng",
    psm=4,
    oem=3,
    extra_config=""
)

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    vlm_model="gpt-5",  # Optional, uses default if None
    api_key="your_api_key"
)

parser = StructuredPDFParser(
    # VLM Engine (pass the initialized engine)
    vlm=vlm_engine,  # or None to disable VLM
    
    # Layout Detection Settings
    layout_model_name="PP-DocLayout_plus-L",
    dpi=200,
    min_score=0.0,
    
    # OCR Engine (pass the initialized engine)
    ocr_engine=ocr_engine,  # or None for default PyTesseract
    
    # Output Settings
    box_separator="\n"
)

# Option 2: Using PaddleOCR for better accuracy
paddle_ocr = PaddleOCREngine(
    device="gpu",  # Use "cpu" if no GPU available
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False
)

parser = StructuredPDFParser(
    ocr_engine=paddle_ocr,
    # ... other settings
)

EnhancedPDFParser

The EnhancedPDFParser extends the StructuredPDFParser with advanced image restoration capabilities using DocRes. This parser is ideal for processing scanned documents, low-quality PDFs, or documents with visual distortions that need enhancement before parsing.

Key Features:

Image Restoration: Uses DocRes for document enhancement before processing
Multiple Restoration Tasks: Supports dewarping, deshadowing, appearance enhancement, deblurring, binarization, and end-to-end restoration
Enhanced Quality: Improves document quality for better OCR and layout detection
All StructuredPDFParser Features: Inherits all capabilities of the base parser
Flexible Configuration: Extensive options for restoration and processing

Basic Usage:

from doctra.parsers.enhanced_pdf_parser import EnhancedPDFParser

# Basic enhanced parser with image restoration
parser = EnhancedPDFParser(
    use_image_restoration=True,
    restoration_task="appearance"  # Default restoration task
)

# Parse document with enhancement
parser.parse("scanned_document.pdf")

Advanced Configuration:

from doctra.engines.ocr import PytesseractOCREngine, PaddleOCREngine

# Initialize OCR engine (PyTesseract or PaddleOCR)
ocr_engine = PytesseractOCREngine(
    lang="eng",
    psm=6,
    oem=3
)

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    vlm_model="gpt-4-vision",  # Optional, uses default if None
    api_key="your_api_key"
)

parser = EnhancedPDFParser(
    # Image Restoration Settings
    use_image_restoration=True,
    restoration_task="dewarping",      # Correct perspective distortion
    restoration_device="cuda",         # Use GPU for faster processing
    restoration_dpi=300,               # Higher DPI for better quality
    
    # VLM Engine (pass the initialized engine)
    vlm=vlm_engine,  # or None to disable VLM
    
    # Layout Detection Settings
    layout_model_name="PP-DocLayout_plus-L",
    dpi=200,
    min_score=0.5,
    
    # OCR Engine (pass the initialized engine)
    ocr_engine=ocr_engine,  # or None for default PyTesseract
)

DocRes Restoration Tasks:

Task	Description	Best For
`appearance`	General appearance enhancement	Most documents (default)
`dewarping`	Correct perspective distortion	Scanned documents with perspective issues
`deshadowing`	Remove shadows and lighting artifacts	Documents with shadow problems
`deblurring`	Reduce blur and improve sharpness	Blurry or low-quality scans
`binarization`	Convert to black and white	Documents needing clean binarization
`end2end`	Complete restoration pipeline	Severely degraded documents

ChartTablePDFParser

The ChartTablePDFParser is a specialized parser focused specifically on extracting charts and tables from PDF documents. It's optimized for scenarios where you only need these specific elements, providing faster processing and more targeted output.

Key Features:

Focused Extraction: Extracts only charts and/or tables
Selective Processing: Choose to extract charts, tables, or both
VLM Integration: Optional conversion to structured data
Organized Output: Separate directories for charts and tables
Progress Tracking: Real-time progress bars for extraction

Basic Usage:

from doctra.parsers.table_chart_extractor import ChartTablePDFParser

# Extract both charts and tables
parser = ChartTablePDFParser(
    extract_charts=True,
    extract_tables=True
)

# Extract only charts
parser = ChartTablePDFParser(
    extract_charts=True,
    extract_tables=False
)

# Parse with custom output directory
parser.parse("document.pdf", output_base_dir="my_outputs")

Advanced Configuration:

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    vlm_model="gpt-5",  # Optional, uses default if None
    api_key="your_api_key"
)

parser = ChartTablePDFParser(
    # Extraction Settings
    extract_charts=True,
    extract_tables=True,
    
    # VLM Engine (pass the initialized engine)
    vlm=vlm_engine,  # or None to disable VLM
    
    # Layout Detection Settings
    layout_model_name="PP-DocLayout_plus-L",
    dpi=200,
    min_score=0.0
)

PaddleOCRVLPDFParser

The PaddleOCRVLPDFParser uses PaddleOCRVL (Vision-Language Model) for end-to-end document parsing. It combines PaddleOCRVL's advanced document understanding capabilities with DocRes image restoration and split table merging, providing a comprehensive solution for complex document processing.

Installation Requirements

Before using PaddleOCRVLPDFParser, install the required dependencies:

pip install -U "paddleocr[doc-parser]"

For Linux systems:

python -m pip install https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl

For Windows systems:

python -m pip install https://xly-devops.cdn.bcebos.com/safetensors-nightly/safetensors-0.6.2.dev0-cp38-abi3-win_amd64.whl

Key Features:

End-to-End Parsing: Uses PaddleOCRVL for complete document understanding in a single pass
Chart Recognition: Automatically extracts and converts charts to structured table format
Document Restoration: Optional DocRes integration for enhanced document quality
Split Table Merging: Automatically detects and merges tables split across pages
Structured Output: Generates Markdown, HTML, and Excel files with tables and charts
Multiple Element Types: Handles headers, text, tables, charts, footnotes, and more

Basic Usage:

from doctra import PaddleOCRVLPDFParser

# Basic parser with default settings
parser = PaddleOCRVLPDFParser(
    use_image_restoration=True,      # Enable DocRes restoration
    use_chart_recognition=True,       # Enable chart recognition
    merge_split_tables=True,          # Enable split table merging
    device="gpu"                      # Use GPU for processing
)

# Parse a PDF document
parser.parse("document.pdf")

Advanced Configuration:

from doctra import PaddleOCRVLPDFParser

parser = PaddleOCRVLPDFParser(
    # DocRes Image Restoration Settings
    use_image_restoration=True,
    restoration_task="appearance",    # Options: appearance, dewarping, deshadowing, deblurring, binarization, end2end
    restoration_device="cuda",        # or "cpu" or None for auto-detect
    restoration_dpi=300,              # DPI for restoration processing
    
    # PaddleOCRVL Settings
    use_chart_recognition=True,       # Enable chart recognition and extraction
    use_doc_orientation_classify=True, # Enable document orientation classification
    use_doc_unwarping=True,           # Enable document unwarping
    use_layout_detection=True,        # Enable layout detection
    device="gpu",                     # "gpu" or "cpu"
    
    # Split Table Merging Settings
    merge_split_tables=True,          # Enable split table detection and merging
    bottom_threshold_ratio=0.20,      # Ratio for "too close to bottom" detection
    top_threshold_ratio=0.15,         # Ratio for "too close to top" detection
    max_gap_ratio=0.25,               # Maximum allowed gap between tables
    column_alignment_tolerance=10.0,  # Pixel tolerance for column alignment
    min_merge_confidence=0.65         # Minimum confidence score for merging
)

# Parse with custom output directory
parser.parse("document.pdf", output_dir="custom_output")

Output Structure:

The parser generates output in outputs/{document_name}/paddleocr_vl_parse/ with:

result.md: Markdown file with all extracted content
result.html: HTML file with formatted output
tables.xlsx: Excel file containing all tables and charts as structured data
tables.html: HTML file with structured tables and charts
enhanced_pages/: Directory with DocRes-enhanced page images (if restoration enabled)
tables/: Directory with merged table images (if split tables detected)

Example Output:

The parser extracts various document elements:

Headers: Document titles and section headers
Text: Paragraphs and body text
Tables: Extracted as HTML and converted to Excel format
Charts: Converted from visual format to structured table data
Footnotes: Vision-based footnote detection
Figure Titles: Captions and figure descriptions

StructuredDOCXParser

The StructuredDOCXParser is a comprehensive parser for Microsoft Word documents (.docx files) that extracts text, tables, images, and structured content while preserving document formatting and order. It supports VLM integration for enhanced content analysis and structured data extraction.

Key Features:

Complete DOCX Support: Extracts text, tables, images, and formatting from Word documents
Document Order Preservation: Maintains the original sequence of elements (paragraphs, tables, images)
VLM Integration: Optional Vision Language Model support for image analysis and table extraction
Multiple Output Formats: Generates Markdown, HTML, and Excel files
Excel Export: Creates structured Excel files with Table of Contents and clickable hyperlinks
Formatting Preservation: Maintains text formatting (bold, italic, etc.) in output
Progress Tracking: Real-time progress bars for VLM processing

Basic Usage:

from doctra.parsers.structured_docx_parser import StructuredDOCXParser

# Basic DOCX parsing
parser = StructuredDOCXParser(
    extract_images=True,
    preserve_formatting=True,
    table_detection=True,
    export_excel=True
)

# Parse DOCX document
parser.parse("document.docx")

Advanced Configuration with VLM:

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",  # or "gemini", "anthropic", "openrouter", "qianfan", "ollama"
    vlm_model="gpt-4-vision",  # Optional, uses default if None
    api_key="your_api_key"
)

parser = StructuredDOCXParser(
    # VLM Engine (pass the initialized engine)
    vlm=vlm_engine,  # or None to disable VLM
    
    # Processing Options
    extract_images=True,
    preserve_formatting=True,
    table_detection=True,
    export_excel=True
)

# Parse with VLM enhancement
parser.parse("document.docx")

Output Structure:

When parsing a DOCX document, the parser creates:

outputs/document_name/
├── document.md          # Markdown version with all content
├── document.html        # HTML version with styling
├── tables.xlsx         # Excel file with extracted tables
│   ├── Table of Contents  # Summary sheet with hyperlinks
│   ├── Table 1         # Individual table sheets
│   ├── Table 2
│   └── ...
└── images/             # Extracted images
    ├── image1.png
    ├── image2.jpg
    └── ...

VLM Integration Features:

When VLM is enabled, the parser:

Analyzes Images: Uses AI to extract structured data from images
Creates Tables: Converts chart images to structured table data
Enhanced Excel Output: Includes VLM-extracted tables in Excel file
Smart Content Display: Shows extracted tables instead of images in Markdown/HTML
Progress Tracking: Shows progress based on number of images processed

CLI Usage:

# Basic DOCX parsing
doctra parse-docx document.docx

# With VLM enhancement
doctra parse-docx document.docx --use-vlm --vlm-provider openai --vlm-api-key your_key

# Custom options
doctra parse-docx document.docx \
  --extract-images \
  --preserve-formatting \
  --table-detection \
  --export-excel

DocResEngine

The DocResEngine provides direct access to DocRes image restoration capabilities. This engine is perfect for standalone image restoration tasks or when you need fine-grained control over the restoration process.

Key Features:

Direct Image Restoration: Process individual images or entire PDFs
Multiple Restoration Tasks: All 6 DocRes restoration tasks available
GPU Acceleration: Automatic CUDA detection and optimization
Flexible Input/Output: Support for various image formats and PDFs
Metadata Extraction: Get detailed information about restoration process

Basic Usage:

from doctra.engines.image_restoration import DocResEngine

# Initialize DocRes engine
docres = DocResEngine(device="cuda")  # or "cpu" or None for auto-detect

# Restore a single image
restored_img, metadata = docres.restore_image(
    image="path/to/image.jpg",
    task="appearance"
)

# Restore entire PDF
enhanced_pdf = docres.restore_pdf(
    pdf_path="document.pdf",
    output_path="enhanced_document.pdf",
    task="appearance"
)

Advanced Usage:

# Initialize with custom settings
docres = DocResEngine(
    device="cuda",                    # Force GPU usage
    use_half_precision=True,         # Use half precision for faster processing
    model_path="custom/model.pth",    # Custom model path (optional)
    mbd_path="custom/mbd.pth"        # Custom MBD model path (optional)
)

# Process multiple images
images = ["doc1.jpg", "doc2.jpg", "doc3.jpg"]
for img_path in images:
    restored_img, metadata = docres.restore_image(
        image=img_path,
        task="dewarping"
    )
    print(f"Processed {img_path}: {metadata}")

# Batch PDF processing
pdfs = ["report1.pdf", "report2.pdf"]
for pdf_path in pdfs:
    output_path = f"enhanced_{os.path.basename(pdf_path)}"
    docres.restore_pdf(
        pdf_path=pdf_path,
        output_path=output_path,
        task="end2end"  # Complete restoration pipeline
    )

Supported Restoration Tasks:

Task	Description	Use Case
`appearance`	General appearance enhancement	Default choice for most documents
`dewarping`	Correct document perspective distortion	Scanned documents with perspective issues
`deshadowing`	Remove shadows and lighting artifacts	Documents with shadow problems
`deblurring`	Reduce blur and improve sharpness	Blurry or low-quality scans
`binarization`	Convert to black and white	Documents needing clean binarization
`end2end`	Complete restoration pipeline	Severely degraded documents

🖥️ Web UI (Gradio)

Doctra provides a comprehensive web interface built with Gradio that makes document processing accessible to non-technical users.

Features:

Drag & Drop Interface: Upload PDFs by dragging and dropping
Multiple Parsers: Choose between full parsing, enhanced parsing, and chart/table extraction
Real-time Processing: See progress as documents are processed
VLM Integration: Configure API keys for AI features
Output Preview: View results directly in the browser
Download Results: Download processed files as ZIP archives

Launch the Web UI:

from doctra.ui.app import launch_ui

# Launch the web interface
launch_ui()

Or from command line:

python gradio_app.py

Web UI Components:

Full Parse Tab: Complete document processing with page navigation
DOCX Parser Tab: Microsoft Word document parsing with VLM integration
Tables & Charts Tab: Specialized extraction with VLM integration
DocRes Tab: Image restoration with before/after comparison
Enhanced Parser Tab: Enhanced parsing with DocRes integration

Command Line Interface

Doctra includes a powerful CLI for batch processing and automation.

Available Commands:

# Full document parsing
doctra parse document.pdf

# DOCX document parsing
doctra parse-docx document.docx

# Enhanced parsing with image restoration
doctra enhance document.pdf --restoration-task appearance

# Extract only charts and tables
doctra extract charts document.pdf
doctra extract tables document.pdf
doctra extract both document.pdf --use-vlm

# Visualize layout detection
doctra visualize document.pdf

# Quick document analysis
doctra analyze document.pdf

# System information
doctra info

CLI Examples:

# Enhanced parsing with custom settings
doctra enhance document.pdf \
  --restoration-task dewarping \
  --restoration-device cuda \
  --use-vlm \
  --vlm-provider openai \
  --vlm-api-key your_key

# Extract charts with VLM
doctra extract charts document.pdf \
  --use-vlm \
  --vlm-provider gemini \
  --vlm-api-key your_key

# Batch processing
doctra parse *.pdf --output-dir results/

🎨 Visualization

Doctra provides powerful visualization capabilities to help you understand how the layout detection works and verify the accuracy of element extraction.

Layout Detection Visualization

The StructuredPDFParser includes a built-in visualization method that displays PDF pages with bounding boxes overlaid on detected elements. This is perfect for:

Debugging: Verify that layout detection is working correctly
Quality Assurance: Check the accuracy of element identification
Documentation: Create visual documentation of extraction results
Analysis: Understand document structure and layout patterns

Basic Visualization:

from doctra.parsers.structured_pdf_parser import StructuredPDFParser

# Initialize parser (OCR engine is optional for visualization)
parser = StructuredPDFParser()

# Display visualization (opens in default image viewer)
parser.display_pages_with_boxes("document.pdf")

Advanced Visualization with Custom Settings:

# Custom visualization configuration
parser.display_pages_with_boxes(
    pdf_path="document.pdf",
    num_pages=5,        # Number of pages to visualize
    cols=3,             # Number of columns in grid
    page_width=600,     # Width of each page in pixels
    spacing=30,         # Spacing between pages
    save_path="layout_visualization.png"  # Save to file instead of displaying
)

Visualization Features:

Color-coded Elements: Each element type (text, table, chart, figure) has a distinct color
Confidence Scores: Shows detection confidence for each element
Grid Layout: Multiple pages displayed in an organized grid
Interactive Legend: Color legend showing all detected element types
High Quality: High-resolution output suitable for documentation
Flexible Output: Display on screen or save to file

Example Output:

The visualization shows:

Blue boxes: Text elements
Red boxes: Tables
Green boxes: Charts
Orange boxes: Figures
Labels: Element type and confidence score (e.g., "table (0.95)")
Page titles: Page number and element count
Summary statistics: Total elements detected by type

Use Cases for Visualization:

Document Analysis: Quickly assess document structure and complexity
Quality Control: Verify extraction accuracy before processing
Debugging: Identify issues with layout detection
Documentation: Create visual reports of extraction results
Training: Help users understand how the system works

Visualization Configuration Options:

Parameter	Default	Description
`num_pages`	3	Number of pages to visualize
`cols`	2	Number of columns in grid layout
`page_width`	800	Width of each page in pixels
`spacing`	40	Spacing between pages in pixels
`save_path`	None	Path to save visualization (if None, displays on screen)

📓 Interactive Notebooks

Notebook	Colab Badge	Description
01_doctra_quick_start		Comprehensive tutorial covering layout detection, content extraction, and multi-format outputs with visual examples

📖 Usage Examples

Example 1: Basic Document Processing

from doctra.parsers.structured_pdf_parser import StructuredPDFParser

# Initialize parser (uses default PyTesseract OCR engine)
parser = StructuredPDFParser()

# Process document
parser.parse("financial_report.pdf")

# Output will be saved to: outputs/financial_report/
# - Extracted text content
# - Cropped images of figures, charts, and tables
# - Markdown file with all content

Example 2: Enhanced Parsing with Image Restoration

from doctra.parsers.enhanced_pdf_parser import EnhancedPDFParser
from doctra.engines.ocr import PytesseractOCREngine

# Initialize OCR engine (optional - defaults to PyTesseract if not provided)
ocr_engine = PytesseractOCREngine(lang="eng", psm=4, oem=3)

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    api_key="your_api_key"
)

# Initialize enhanced parser with image restoration
parser = EnhancedPDFParser(
    use_image_restoration=True,
    restoration_task="dewarping",  # Correct perspective distortion
    restoration_device="cuda",    # Use GPU for faster processing
    ocr_engine=ocr_engine,        # Pass OCR engine instance
    vlm=vlm_engine                # Pass VLM engine instance
)

# Process scanned document with enhancement
parser.parse("scanned_document.pdf")

# Output will include:
# - Enhanced PDF with restored images
# - All standard parsing outputs
# - Improved OCR accuracy due to restoration

Example 2b: Using PaddleOCR for Better Accuracy

from doctra.parsers.structured_pdf_parser import StructuredPDFParser
from doctra.engines.ocr import PaddleOCREngine

# Initialize PaddleOCR engine with custom settings
paddle_ocr = PaddleOCREngine(
    device="gpu",  # Use "cpu" if no GPU available
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False
)

# Create parser with PaddleOCR engine
parser = StructuredPDFParser(
    ocr_engine=paddle_ocr
)

# Process document with PaddleOCR
parser.parse("complex_document.pdf")

# PaddleOCR provides:
# - Higher accuracy for complex documents
# - Better performance on GPU
# - Automatic model management

Example 3: Direct Image Restoration

from doctra.engines.image_restoration import DocResEngine

# Initialize DocRes engine
docres = DocResEngine(device="cuda")

# Restore individual images
restored_img, metadata = docres.restore_image(
    image="blurry_document.jpg",
    task="deblurring"
)

# Restore entire PDF
docres.restore_pdf(
    pdf_path="low_quality.pdf",
    output_path="enhanced.pdf",
    task="appearance"
)

Example 4: DOCX Document Parsing

from doctra.parsers.structured_docx_parser import StructuredDOCXParser

# Basic DOCX parsing
parser = StructuredDOCXParser(
    extract_images=True,
    preserve_formatting=True,
    table_detection=True,
    export_excel=True
)

# Parse Word document
parser.parse("report.docx")

# Output will include:
# - Markdown file with all content
# - HTML file with styling
# - Excel file with extracted tables
# - Extracted images in organized folders

Example 5: DOCX Parsing with VLM Enhancement

from doctra.parsers.structured_docx_parser import StructuredDOCXParser

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    vlm_model="gpt-4-vision",  # Optional, uses default if None
    api_key="your_api_key"
)

# DOCX parsing with VLM for enhanced analysis
parser = StructuredDOCXParser(
    vlm=vlm_engine,  # Pass VLM engine instance
    extract_images=True,
    preserve_formatting=True,
    table_detection=True,
    export_excel=True
)

# Parse with AI enhancement
parser.parse("financial_report.docx")

# Output will include:
# - All standard outputs
# - VLM-extracted tables from images
# - Enhanced Excel with Table of Contents
# - Smart content display (tables instead of images)

Example 6: PaddleOCRVL End-to-End Parsing

from doctra import PaddleOCRVLPDFParser

# Initialize PaddleOCRVL parser with all features enabled
parser = PaddleOCRVLPDFParser(
    use_image_restoration=True,      # Enable DocRes restoration
    restoration_task="appearance",    # Use appearance enhancement
    use_chart_recognition=True,       # Enable chart recognition
    merge_split_tables=True,          # Enable split table merging
    device="gpu"                      # Use GPU for faster processing
)

# Parse document - automatically handles all content types
parser.parse("financial_report.pdf")

# Output will be in: outputs/financial_report/paddleocr_vl_parse/
# - result.md: All content in Markdown
# - result.html: Formatted HTML output
# - tables.xlsx: All tables and charts in Excel format
# - tables.html: Structured tables and charts

Example 7: Chart and Table Extraction with VLM

from doctra.parsers.table_chart_extractor import ChartTablePDFParser

# Initialize VLM engine
from doctra.engines.vlm.service import VLMStructuredExtractor

vlm_engine = VLMStructuredExtractor(
    vlm_provider="openai",
    api_key="your_api_key"
)

# Initialize parser with VLM
parser = ChartTablePDFParser(
    extract_charts=True,
    extract_tables=True,
    vlm=vlm_engine  # Pass VLM engine instance
)

# Process document
parser.parse("data_report.pdf", output_base_dir="extracted_data")

# Output will include:
# - Cropped chart and table images
# - Structured data in Excel format
# - Markdown tables with extracted data

Example 8: Web UI Usage

from doctra.ui.app import launch_ui

# Launch the web interface
launch_ui()

# Or build the interface programmatically
from doctra.ui.app import build_demo
demo = build_demo()
demo.launch(share=True)  # Share publicly

Example 9: Command Line Usage

# DOCX parsing with VLM
doctra parse-docx document.docx \
  --use-vlm \
  --vlm-provider openai \
  --vlm-api-key your_key \
  --extract-images \
  --export-excel

# Enhanced parsing with custom settings
doctra enhance document.pdf \
  --restoration-task dewarping \
  --restoration-device cuda \
  --use-vlm \
  --vlm-provider openai \
  --vlm-api-key your_key

# Extract charts with VLM
doctra extract charts document.pdf \
  --use-vlm \
  --vlm-provider gemini \
  --vlm-api-key your_key

# Batch processing
doctra parse *.pdf --output-dir results/

Example 10: Layout Visualization

from doctra.parsers.structured_pdf_parser import StructuredPDFParser

# Initialize parser (OCR engine not needed for visualization)
parser = StructuredPDFParser()

# Create a comprehensive visualization
parser.display_pages_with_boxes(
    pdf_path="research_paper.pdf",
    num_pages=6,        # Visualize first 6 pages
    cols=2,             # 2 columns layout
    page_width=700,     # Larger pages for better detail
    spacing=50,         # More spacing between pages
    save_path="research_paper_layout.png"  # Save for documentation
)

# For quick preview (displays on screen)
parser.display_pages_with_boxes("document.pdf")

✨ Features

🔍 Layout Detection

Advanced document layout analysis using PaddleOCR
Accurate identification of text, tables, charts, and figures
Configurable confidence thresholds

📝 OCR Processing

Dual OCR Engine Support: Choose between PyTesseract (default) or PaddleOCR PP-OCRv5_server
Dependency Injection Pattern: Initialize OCR engines externally and pass them to parsers for clearer API
PaddleOCR PP-OCRv5_server: Advanced model from PaddleOCR 3.0 with superior accuracy
PyTesseract: Traditional OCR with extensive language support and fine-grained control
Reusable Engines: Create OCR engine instances once and reuse across multiple parsers
Support for multiple languages (PyTesseract)
GPU acceleration for PaddleOCR
Configurable OCR parameters for both engines

🧠 PaddleOCRVL End-to-End Parsing

Vision-Language Model: Advanced document understanding using PaddleOCRVL
Complete Document Parsing: Single-pass extraction of all content types
Chart Recognition: Automatic chart detection and conversion to structured tables
Multi-Element Support: Handles headers, text, tables, charts, footnotes, and figure titles
Integrated Restoration: Optional DocRes image restoration for enhanced quality
Split Table Merging: Automatic detection and merging of tables across pages
Structured Output: Generates Excel files with both tables and charts

🖼️ Visual Element Extraction

Automatic cropping and saving of figures, charts, and tables
Organized output directory structure
High-resolution image preservation

🔧 Image Restoration (DocRes)

6 Restoration Tasks: Dewarping, deshadowing, appearance enhancement, deblurring, binarization, and end-to-end restoration
GPU Acceleration: Automatic CUDA detection and optimization
Enhanced Quality: Improves document quality for better OCR and layout detection
Flexible Processing: Standalone image restoration or integrated with parsing

🤖 VLM Integration

Dependency Injection Pattern: Initialize VLM engines externally and pass them to parsers for clearer API
Vision Language Model Support: Structured data extraction from visual elements
Multiple Provider Options: OpenAI, Gemini, Anthropic, OpenRouter, Qianfan, Ollama
Reusable Engines: Create VLM engine instances once and reuse across multiple parsers
Automatic Conversion: Charts and tables converted to structured formats (Excel, HTML, JSON)

📊 Multiple Output Formats

Markdown: Human-readable document with embedded images and tables
Excel: Structured data in spreadsheet format
JSON: Programmatically accessible structured data
HTML: Interactive web-ready documents
Images: High-quality cropped visual elements

🖥️ User Interfaces

Web UI: Gradio-based interface with drag & drop functionality
Command Line: Powerful CLI for batch processing and automation
Multiple Tabs: Full parsing, DOCX parsing, enhanced parsing, chart/table extraction, and image restoration

⚙️ Flexible Configuration

Extensive customization options
Performance tuning parameters
Output format selection
Device selection (CPU/GPU)

🙏 Acknowledgments

Doctra builds upon several excellent open-source projects:

PaddleOCR - Advanced document layout detection and OCR capabilities
DocRes - State-of-the-art document image restoration model
Outlines - Structured output generation for LLMs

We thank the developers and contributors of these projects for their valuable work that makes Doctra possible.

Keywords

FAQs

What is doctra?

Is doctra well maintained?

Did you know?

Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.

Install