Docling PP-OCRv6 Plugin

A Docling OCR plugin that runs PaddlePaddle's PP-OCRv6 (medium) detection and recognition models locally through RapidOCR and onnxruntime. No external service is needed.

GitHub Repository: DCC-BS/docling-pp-ocrv6

Features

• Local OCR: Runs inside the docling worker, on the GPU when docling's accelerator is CUDA and onnxruntime-gpu is installed.
• European Languages: Defaults to a German-led European language set.
• Word Boxes: return_word_box adds each word OCR reads, with its box, to the page's word cells.
• Whole-Page Reading: whole_page reads each page as one picture instead of layout crops, so lines are not cut at layout boxes.
• Engine Key: Registers under the "pp-ocrv6" engine key.

Installation

Pick exactly one onnxruntime extra:

• CPU: pip install "docling-pp-ocrv6[cpu]"
• CUDA: pip install "docling-pp-ocrv6[gpu]"

Usage

from docling_pp_ocrv6 import PPOCRv6Options

pipeline_options.ocr_options = PPOCRv6Options(return_word_box=True)

With docling-serve: { "options": { "do_ocr": true, "ocr_engine": "pp-ocrv6" } }. Options can also be set with PPOCRV6_* environment variables (e.g. PPOCRV6_LANG, PPOCRV6_WHOLE_PAGE).

Environment Variables

Variable	Description	Default
PPOCRV6_LANG	Comma-separated languages	German-led European set
PPOCRV6_TEXT_SCORE	Minimum recognition score	0.5
PPOCRV6_USE_DET / _USE_CLS / _USE_REC	Run detection / angle classification / recognition	true
PPOCRV6_WHOLE_PAGE	Read every page as one picture	false
PPOCRV6_RETURN_WORD_BOX	Return a box per OCR'd word	false
PPOCRV6_DET_REPO / _REC_REPO	HuggingFace repos of the ONNX models	PaddlePaddle/PP-OCRv6_medium_{det,rec}_onnx
PPOCRV6_DET_MODEL_PATH / _REC_MODEL_PATH / _REC_KEYS_PATH / _CLS_MODEL_PATH	Local model files instead of the repos	auto