This library orchestrates document layout analysis and extraction for RAG, targeting researchers and developers building Document AI pipelines. It offers a unified framework for training, evaluating, and inferencing models, simplifying complex document understanding tasks.

How It Works

Deepdoctection integrates multiple state-of-the-art models for layout analysis, OCR, and document classification. It leverages PyTorch with Detectron2 and Transformers, or TensorFlow with Tensorpack, for core vision tasks. For OCR, it supports Tesseract, DocTr, and AWS Textract. Document and token classification are handled by LayoutLM family models, LiLT, and BERT-style architectures, incorporating features like sliding windows. Additional utilities include text mining for PDFs, language detection, and image deskewing.

Quick Start & Requirements

• Install via pip: pip install deepdoctection (with [pt] or [tf] for full features).
• Requires Python >= 3.9.
• PyTorch (>= 2.2) or TensorFlow (>= 2.11, < 2.16) is necessary. TensorFlow support is deprecated from Python 3.11.
• GPU recommended for fine-tuning.
• Linux or macOS; Windows is not supported but a Dockerfile is available.
• See introduction notebook and Hugging Face Space Demo.

Highlighted Details

• Supports Detectron2, Transformers, Tensorpack, Tesseract, DocTr, AWS Textract, LayoutLM, LiLT.
• Offers fine-tuning and evaluation tools.
• Includes tutorials and a Hugging Face Space demo.
• Handles native PDFs via pdfplumber.

Maintenance & Community

• Active development with recent v0.43 release.
• Community support via GitHub stars and recommendations.

Licensing & Compatibility

• Apache 2.0 License.
• Permissive license suitable for commercial use and closed-source linking.

Limitations & Caveats

Windows is not officially supported, though a Docker solution exists. TensorFlow support is being phased out for newer Python versions.