Docling PP-Doc-Layout Plugin
A Docling plugin that provides high-accuracy document layout detection using the PaddlePaddle PP-DocLayoutV3 model.
GitHub Repository: DCC-BS/docling-pp-doc-layout
Features
- High Accuracy: Utilizes the RT-DETR instance segmentation framework.
- Polygon Support: Gracefully flattens complex polygon masks into Docling-compatible bounding boxes.
- List Detection: The model has no list class, so the plugin detects list items from bullets and enumerator sequences after OCR (
PP_DOC_LAYOUT_LIST_DETECTION=rules, default), optionally with docling's Heron model as a second opinion (heron). - Keeps Handwriting: The default confidence threshold is 0.3 (
PP_DOC_LAYOUT_CONFIDENCE_THRESHOLD), so handwriting and text in photos are kept. - Scalability: Supports configurable batch sizing to optimize GPU VRAM usage and prevent OOM errors.
- Auto-Registration: Automatically registers itself as a layout engine upon installation.
Installation
- Using uv (recommended):
uv add docling-pp-doc-layout - Using pip:
pip install docling-pp-doc-layout
Usage
Integrate into the Docling Python SDK by configuring PdfPipelineOptions:
python
from docling_pp_doc_layout.options import PPDocLayoutV3Options
pipeline_options.layout_options = PPDocLayoutV3Options(batch_size=8)Environment Variables
| Variable | Description | Default |
|---|---|---|
PP_DOC_LAYOUT_MODEL_NAME | HuggingFace repo or local path of the model | PaddlePaddle/PP-DocLayoutV3_safetensors |
PP_DOC_LAYOUT_CONFIDENCE_THRESHOLD | Minimum detection confidence | 0.3 |
PP_DOC_LAYOUT_BATCH_SIZE | Pages per inference batch | 8 |
PP_DOC_LAYOUT_LIST_DETECTION | rules, heron or off | rules |
PP_DOC_LAYOUT_CREATE_ORPHAN_CLUSTERS | Create clusters for orphaned elements | true |
PP_DOC_LAYOUT_KEEP_EMPTY_CLUSTERS | Keep empty clusters | false |
PP_DOC_LAYOUT_SKIP_CELL_ASSIGNMENT | Skip table-cell assignment | false |