Offline PDF Parser using Docling

This project provides a complete offline PDF processing solution using Docling, capable of extracting text, tables, and images from PDF documents without requiring an internet connection.

Features

• Completely Offline: All models are pre-downloaded and stored locally
• Comprehensive Extraction: Extracts text, tables, images, and document structure
• Multiple Output Formats: JSON, Markdown, and plain text
• GPU Acceleration: Uses CUDA for faster processing when available
• Table Recognition: Advanced table structure recognition and extraction
• OCR Support: Optical Character Recognition for scanned documents

Setup

1. Environment Setup

# Create conda environment
conda create -n docling python=3.12 -y
conda activate docling

# Install docling
pip install docling

2. Download Models (One-time setup)

# Download all required models for offline usage
docling-tools models download

This will download models to ~/.cache/docling/models/ (approximately 1.2GB).

Usage

Basic Usage

# Activate environment
conda activate docling

# Run the parser
python final_offline_parser.py

Files

• final_offline_parser.py - Main parser script (recommended)
• offline_pdf_parser.py - Original parser script
• fixed_offline_parser.py - Fixed version
• test_offline.py - Test script to verify setup
• debug_document_structure.py - Debug script for document structure analysis

Output

The parser creates an output/ directory with:

• companies_house_document_content.json - Complete structured data
• companies_house_document_content.md - Markdown formatted output
• companies_house_document_text.txt - Plain text output

Example Results

For the companies_house_document.pdf:

• Pages processed: 32
• Tables found: 20
• Figures found: 3
• Text extracted: 76,975 characters
• Processing time: ~4 minutes (with GPU acceleration)

Technical Details

Offline Configuration

The parser is configured to run completely offline:

pipeline_options = PdfPipelineOptions(
    artifacts_path=artifacts_path,
    enable_remote_services=False,  # No internet required
    do_table_structure=True,       # Table recognition
    do_ocr=True,                   # OCR processing
    do_chunking=True,              # Document chunking
)

Model Requirements

• Layout Model: Document layout analysis
• Tableformer Model: Table structure recognition
• Picture Classifier: Image classification
• Code Formula Model: Mathematical formula recognition
• EasyOCR Models: Text recognition

Troubleshooting

Common Issues

1. Models not found: Run docling-tools models download
2. CUDA errors: Ensure NVIDIA drivers and CUDA are properly installed
3. Memory issues: Reduce batch size or use CPU-only mode

Testing Setup

python test_offline.py

This will verify that all models are downloaded and the environment is properly configured.

Performance

• GPU Processing: ~4 minutes for 32-page document
• CPU Processing: ~8-10 minutes for 32-page document
• Memory Usage: ~2-4GB RAM during processing
• Storage: ~1.2GB for all models

Model Structure

For detailed information about all model files and folder structure, see MODEL_STRUCTURE.md.

License

This project uses Docling, which is licensed under the Apache License 2.0.