PaddlePaddle/PaddleOCR: Turn any PDF or image document into structured data for your AI. A powerful, lightweight OCR toolkit that bridges the gap between images/PDFs and LLMs. Supports 100+ languages.

Turn any PDF or image document into structured data for your AI. A powerful, lightweight OCR toolkit that bridges the gap between images/PDFs and LLMs. Supports 100+ languages. - PaddlePaddle/Paddl...

Summary

PaddleOCR is an Apache-2.0 Python toolkit (Python 3.8–3.12, Linux/Windows/macOS) that converts PDFs and images into structured Markdown or JSON, with text recognition support for 100+ languages. It bundles three main pipelines: PP-OCRv6 for fast scene-text OCR, PP-StructureV3 for layout-aware parsing with cell-level coordinates, and PaddleOCR-VL, a 0.9B vision-language model for document understanding.

What it is useful for

Engineers building RAG, document Q&A, or agentic systems usually need clean text and structured tables from messy inputs. PaddleOCR covers that with three documented pipelines. PP-OCRv6 reads text from natural scenes and supports 50 languages in a single unified model, useful when a document mixes Chinese, English, and Latin-script languages without swapping models. PP-StructureV3 returns Markdown or JSON with text, table cell, and formula coordinates, helpful when downstream code needs to map output back to source positions. PaddleOCR-VL and HPD-Parsing handle harder cases such as charts, seals, ancient documents, and cross-page table joining, with Markdown output that can feed an LLM context directly. The README documents integrations with Dify, RAGFlow, Pathway, and Cherry Studio for RAG and agent workflows, plus an ONNX export path and an official browser inference SDK called PaddleOCR.js.

How engineers can use it

The README does not include runnable install commands directly. It points engineers to dedicated documentation pages for each pipeline and to the official website, which exposes an interactive experience center for trying models without local setup. For local work, the README links to separate guides covering PP-OCR, PaddleOCR-VL, and PP-StructureV3, plus installation notes covering CPU, NVIDIA GPU, Intel CPU, Kunlunxin XPU, and other AI accelerators. Models are published on HuggingFace and ModelScope. Server-side usage supports an OpenAI-compatible endpoint and a custom vLLM runtime for HPD-Parsing, while client SDKs cover C++, C#, and Java. Engineers should follow the docs linked from the README before installing, since the README itself only lists documentation entry points. A noted dependency detail from the 3.2.0 release: core and optional dependencies were separated, so only minimal packages are needed for basic text recognition, with extras required for document parsing and information extraction.