PaddleOCR converts PDF documents and images into structured Markdown or JSON for use in retrieval-augmented generation and agent pipelines. It combines classic OCR pipelines for scene text with vision-language models for whole-document parsing, and it runs on CPU, NVIDIA GPU, Kunlunxin XPU, and other accelerators across Linux, Windows, and macOS.
Document parsing is handled by two families: the PaddleOCR-VL series (a 0.9B-parameter vision-language model that pairs a NaViT-style dynamic-resolution encoder with the ERNIE-4.5-0.3B language model) and PP-StructureV3, a pipeline that adds fine-grained coordinates for table cells and text blocks. Scene text recognition uses the PP-OCR series; PP-OCRv6 covers 50 languages in a single model and ships in tiny (1.5M), small (7.7M), and medium (34.5M) parameter tiers for edge, mobile, and server deployment.
| Model | Task | Reported figure |
|---|---|---|
| PaddleOCR-VL-1.6 | Document parsing | 96.3% on OmniDocBench v1.6 |
| PaddleOCR-VL-1.5 | Document parsing | 94.5% on OmniDocBench, 111 languages |
| PP-OCRv6 medium | Scene text | +4.6% detection and +5.1% recognition over PP-OCRv5_server |
| HPD-Parsing | High-throughput parsing | Peak 4,752 tokens/s with hierarchical parallel decoding |
Features
- Document to Markdown/JSON: parses text, tables, formulas, charts, seals, and layout into structured output, with DOCX export for parsed results
- Multilingual scene OCR: text spotting across IDs, street views, books, and industrial components, with recognition for 100+ languages
- Office document conversion: converts Word, Excel, and PowerPoint files into Markdown
- Document translation: the PP-DocTranslation pipeline translates parsed documents
- Long-document handling: cross-page table merging and hierarchical heading identification
- Inference backends: switches between Paddle static graph, Paddle dynamic graph, and Transformers, with 20 models supporting the Transformers backend
- Browser inference: PaddleOCR.js runs PP-OCRv5 directly in the browser
- Serving options: HPD-Parsing supports OpenAI-compatible serving and local inference through a customised vLLM runtime
- Ecosystem integrations: used within Dify, RAGFlow, Pathway, and Cherry Studio
- Data pipeline: tooling for building datasets to fine-tune large language models
Integrates with
Cherry Studio
Cross-platform desktop client for interacting with multiple LLM providers
Dify
Open-source LLM app development platform combining a visual workflow canvas, RAG pipelines, agents, model management, and observability
RAGFlow
RAGFlow is a leading open-source Retrieval-Augmented Generation (RAG) engine that fuses cutting-edge RAG with Agent capabilities to create a superior context layer for LLMs
vLLM
LLM inference and serving library using PagedAttention and continuous batching, with an OpenAI-compatible API server
