Problem solved
Convert PDF documents and images into structured, LLM-ready data (JSON/Markdown) with high accuracy. Provide multilingual OCR (100+ languages) and document layout parsing for RAG and agentic AI applications without relying on closed-source commercial solutions.
How it works
PaddleOCR comprises three main components: PP-OCRv6 (text detection and recognition engine supporting 50 languages in a unified model), PaddleOCR-VL-1.6 (0.9B lightweight vision-language model for document parsing achieving 96.3% accuracy on OmniDocBench), and PP-StructureV3 (structure-aware PDF/image-to-Markdown/JSON converter with fine-grained coordinate extraction). Built on PaddlePaddle deep learning framework (Python/C++), it supports inference on NVIDIA GPU, Intel CPU, Kunlunxin XPU, and other AI accelerators. Output formats include Markdown and JSON with table cell and text coordinates.
Chinese specificity
Developed by Baidu's PaddlePaddle team; PaddlePaddle is Baidu's open-source deep learning framework widely adopted in the Chinese AI ecosystem. Integration with Kunlunxin XPU (Chinese AI accelerator) is explicitly supported. No mandatory Chinese standard compliance cited.
Western equivalent
Tesseract (open-source OCR engine), EasyOCR (Python library), Docling (IBM, document parsing), PyMuPDF (PDF extraction)