Getting Started — Overview

PaddleOCR is an open-source, production-grade OCR (Optical Character Recognition) toolkit and document AI engine developed by Baidu's PaddlePaddle team. PaddleOCR is designed to bridge the gap between raw image/PDF documents and structured, LLM-consumable data — making it the preferred document parser for AI pipelines in 2025–2026.

PaddleOCR provides a complete document intelligence stack organized into modular pipelines. Each pipeline is purpose-built for a specific document processing task:

• PP-OCRv6 — Universal text detection and recognition (Scene OCR). Supports 50 languages in a single model with three model tiers for different hardware constraints.
• PP-StructureV3 — Complete document layout analysis. Converts PDFs and complex images to structured Markdown or JSON with table cell coordinates, formula positions, and reading-order data.
• PaddleOCR-VL — Vision-language model for document parsing. Uses PaddleOCR-VL-1.6 (0.9B parameters) to achieve 96.33% accuracy on OmniDocBench v1.6 — SOTA for open-source document AI.
• PP-ChatOCRv4 — Conversational document understanding. Chat with your PDFs and images using natural language queries backed by PaddleOCR's extraction pipeline.
• HPD-Parsing — High-throughput document parsing with hierarchical parallel decoding. Achieves 4,752 tokens/second peak throughput for large-scale document processing.

PaddleOCR is trusted by 6,000+ open-source repositories and integrated into the world's most popular RAG and AI agent platforms including Dify, RAGFlow, Pathway, and Cherry Studio.

Installation

PaddleOCR is a Python package distributed via PyPI. It requires Python 3.8 through 3.12 and the PaddlePaddle deep learning framework. PaddleOCR runs on Windows, macOS, and Linux on CPU and GPU.

System Requirements

• Python: 3.8 ≤ version ≤ 3.12 (Python 3.13+ not yet supported)
• OS: Linux (Ubuntu 18.04+), macOS (10.15+), Windows 10/11
• Hardware: CPU (any x86-64 or ARM64), NVIDIA GPU (CUDA 11.8+), Intel CPU with OpenVINO, Apple Silicon (M1/M2/M3/M4)
• RAM: Minimum 4 GB (8 GB recommended for medium model)

Step-by-Step Installation

# Install CPU version (all platforms)
$ pip install paddlepaddle

# For NVIDIA GPU (CUDA 11.8+)
$ pip install paddlepaddle-gpu

# For NVIDIA GPU (CUDA 12.x)
$ pip install paddlepaddle-gpu==3.0.0.post120 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html

$ pip install paddleocr

# Install latest dev version from source
$ git clone https://github.com/PaddlePaddle/PaddleOCR.git
$ cd PaddleOCR && pip install -e .

$ python -c "import paddleocr; print('PaddleOCR', paddleocr.__version__, 'ready')"
# Expected output: PaddleOCR 3.7.0 ready

Install via Docker

# Pull the latest PaddleOCR Docker image
$ docker pull paddlecloud/paddleocr:latest

# Run interactively
$ docker run -it paddlecloud/paddleocr:latest bash

# Run with GPU support (requires nvidia-docker)
$ docker run --gpus all -it paddlecloud/paddleocr:gpu bash

Quick Start Guide

The fastest way to extract text from an image using PaddleOCR is the PaddleOCR class with the predict() method. PaddleOCR automatically handles model loading, text detection, and recognition in a single call.

Basic Text Extraction

from paddleocr import PaddleOCR

# Initialize — downloads models on first run (~150 MB)
ocr = PaddleOCR(lang='en')

# Run OCR on an image
result = ocr.predict('image.jpg')

# Access extracted text
for page in result:
    for text in page['rec_texts']:
        print(text)  # Prints each text block

Get Text with Bounding Boxes

from paddleocr import PaddleOCR

ocr = PaddleOCR(lang='en')
result = ocr.predict('invoice.jpg')

for page in result:
    texts  = page['rec_texts']
    boxes  = page['rec_boxes']   # [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]
    scores = page['rec_scores']  # confidence 0.0–1.0
    
    for text, box, score in zip(texts, boxes, scores):
        print(f'Text: {text}')
        print(f'Box: {box}')
        print(f'Confidence: {score:.3f}')

PDF to Markdown (PP-StructureV3)

from paddleocr import PPStructureV3

# Initialize PP-StructureV3 — full document parser
pipeline = PPStructureV3()

# Parse a PDF — automatically handles multi-page
results = pipeline.predict('document.pdf')

# Save each page as Markdown
for i, page in enumerate(results):
    md = page.get('markdown', '')
    with open(f'page_{i+1}.md', 'w', encoding='utf-8') as f:
        f.write(md)
    print(f'Page {i+1} saved')

PP-OCRv6 — Universal Scene OCR

PP-OCRv6 is PaddleOCR's flagship text detection and recognition pipeline, released with PaddleOCR v3.7.0 in June 2026. It achieves state-of-the-art accuracy on real-world scene text benchmarks, surpassing much larger vision-language models with only 34.5M parameters in the medium tier.

PP-OCRv6 Model Tiers

PP-OCRv6 Key Improvements

• +4.6% detection accuracy over PP-OCRv5_server on standard benchmarks
• +5.1% recognition accuracy — surpassing Qwen3-VL-235B and GPT-5.5 with 34.5M parameters
• 50 languages unified — Chinese, English, Japanese, and 46 Latin-script languages in one model
• 5.2× CPU speedup via OpenVINO backend on Intel hardware
• 6.1× Apple Silicon speedup on M4 (tiny model)
• Specialized improvements for digital displays, dot-matrix characters, tire prints, and industrial text

PP-OCRv5 vs PP-OCRv6 Architecture & Upgrade Guide

PP-OCRv5 was the previous stable release of PaddleOCR (v3.5–v3.6). While PP-OCRv5 remains fully supported for backward compatibility, upgrading to PP-OCRv6 is strongly recommended for all new and existing deployments:

PP-OCRv6 Configuration Options

lang str default: 'en'

Language for recognition. Options include 'en', 'ch', 'japan', 'korean', 'arabic', 'latin', and 80+ more. Use 'ch' for a unified Chinese+English model.

use_gpu bool default: False

Enable GPU acceleration. Set to True when using paddlepaddle-gpu. Dramatically increases throughput for batch processing.

use_angle_cls bool default: False

Enable text angle classification. Set to True for documents with rotated or upside-down text (e.g., scanned documents at various orientations).

det_db_thresh float default: 0.3

Detection binarization threshold. Lower values detect more text regions (may include noise). Higher values are more selective. Range: 0.0–1.0.

det_db_box_thresh float default: 0.6

Minimum confidence threshold for text boxes. Boxes below this threshold are discarded. Increase to 0.8+ for high-precision mode.

rec_batch_num int default: 6

Number of text regions processed per recognition batch. Increase for GPU batching (e.g., 32–64) to maximize GPU utilization during batch processing.

PP-StructureV3 — Document Parsing Pipeline

PP-StructureV3 is PaddleOCR's document layout analysis and structured conversion pipeline. Unlike the base OCR pipeline which extracts flat text, PP-StructureV3 understands the full structure of a document — including reading order, table layout, formula positions, chart locations, and heading hierarchy — and outputs clean, semantically structured Markdown or JSON.

What PP-StructureV3 Can Parse

• Multi-page PDF documents (native PDF and scanned PDFs)
• Images with complex layouts (multi-column academic papers, reports)
• Tables — extracts HTML, cell content, and row/column coordinates
• Mathematical formulas — detected and preserved as LaTeX
• Charts and figures — regions detected and described
• Seals and stamps
• Multi-page TIFF files (v3.6.0+)
• Office documents: Word, Excel, PowerPoint → Markdown (v3.5.0+)

PP-StructureV3 Usage

from paddleocr import PPStructureV3

pipeline = PPStructureV3(
    lang='en',       # Language for text recognition
    use_gpu=False,   # Set True for GPU
)

# Process a PDF or image
results = pipeline.predict('annual_report.pdf')

for page_idx, page in enumerate(results):
    print(f'=== Page {page_idx + 1} ===')
    
    # Get full structured Markdown
    print(page.get('markdown', ''))
    
    # Access individual content blocks
    for block in page.get('blocks', []):
        block_type = block['type']  # 'text', 'table', 'formula', 'figure'
        
        if block_type == 'table':
            print('HTML:', block['html'])
        elif block_type == 'formula':
            print('LaTeX:', block['latex'])

PP-StructureV3 Output JSON Schema

# Each page in results is a dict with this structure:
{
  "page_id": 0,
  "markdown": "# Document Title\n\n...",
  "blocks": [
    {
      "type": "title",  # title|text|table|formula|figure|seal
      "bbox": [x1, y1, x2, y2],
      "text": "..."
    },
    {
      "type": "table",
      "bbox": [x1, y1, x2, y2],
      "html": "<table>...</table>",
      "cells": [...]  # Cell-level coordinates
    }
  ]
}

PaddleOCR-VL — Vision-Language Document AI

PaddleOCR-VL is a lightweight vision-language model (VLM) series specifically optimized for document parsing. Unlike the pipeline-based PP-StructureV3, PaddleOCR-VL processes an entire document page holistically — understanding layout, text, tables, and visual elements together as a vision-language model.

The latest version, PaddleOCR-VL-1.6 (0.9B parameters), achieves a 96.33% score on OmniDocBench v1.6 — the highest accuracy of any open-source document parsing model as of June 2026. It surpasses both commercial closed-source solutions and much larger models in text, formula, and table recognition tasks.

Key Capabilities

• 96.33% accuracy on OmniDocBench v1.6 (current SOTA)
• Table, ancient Chinese document, and rare character recognition
• Seal recognition and text spotting across diverse formats
• Chart and diagram understanding
• Structured output in Markdown and JSON formats
• Zero-cost migration from VL-1.5 — fully compatible model architecture

PaddleOCR-VL Usage

# PaddleOCR-VL available on HuggingFace
from transformers import AutoProcessor, AutoModelForVision2Seq
from PIL import Image

# Load PaddleOCR-VL-1.6 from HuggingFace Hub
processor = AutoProcessor.from_pretrained('PaddlePaddle/PaddleOCR-VL-1.6')
model = AutoModelForVision2Seq.from_pretrained('PaddlePaddle/PaddleOCR-VL-1.6')

# Parse a document image
image = Image.open('complex_document.jpg')
prompt = "Parse this document and output as Markdown."

inputs = processor(
    text=prompt,
    images=image,
    return_tensors='pt'
)
outputs = model.generate(**inputs, max_new_tokens=2048)
result = processor.decode(outputs[0], skip_special_tokens=True)
print(result)

Advanced Usage & Performance Tuning

PaddleOCR provides several configuration options to maximize performance for production workloads. This section covers hardware-specific optimizations, batching strategies, and model selection guidance.

OpenVINO Backend (Intel CPU — 5.2× Speedup)

from paddleocr import PaddleOCR

# Enable MKL-DNN (OpenVINO) for Intel CPU — 5.2× speedup
ocr = PaddleOCR(
    lang='en',
    use_gpu=False,
    enable_mkldnn=True,    # Enable MKL-DNN optimization
    cpu_threads=8,         # Number of CPU threads
)

# Tip: Keep ocr instance across calls — model init is expensive
result = ocr.predict('image.jpg')

GPU Batch Inference

import glob
from paddleocr import PaddleOCR

# Initialize once with large batch size for GPU
ocr = PaddleOCR(
    lang='en',
    use_gpu=True,
    rec_batch_num=32,       # Larger batch = better GPU utilization
    det_limit_side_len=960, # Limit resize for speed vs accuracy tradeoff
)

images = glob.glob('docs/*.jpg')
results = []
for img in images:
    results.append(ocr.predict(img))
    
print(f'Processed {len(results)} images')

RAG & LLM Pipeline Integration

PaddleOCR is the most widely used document parser in open-source RAG (Retrieval-Augmented Generation) pipelines. It bridges the gap between raw document files (PDFs, images, scanned documents) and the structured text chunks needed by vector databases and LLMs.

PaddleOCR → LangChain → Vector Store

from paddleocr import PPStructureV3
from langchain.text_splitter import MarkdownTextSplitter
from langchain.vectorstores import Chroma

# Step 1: Parse PDF with PaddleOCR → Markdown
pipeline = PPStructureV3()
results = pipeline.predict('knowledge_base.pdf')
full_markdown = '\n\n'.join(p.get('markdown', '') for p in results)

# Step 2: Split into LLM-friendly chunks
splitter = MarkdownTextSplitter(chunk_size=1000, chunk_overlap=100)
chunks = splitter.split_text(full_markdown)

# Step 3: Embed and index in a vector store
from langchain.embeddings import OpenAIEmbeddings
vectorstore = Chroma.from_texts(chunks, OpenAIEmbeddings())

# Step 4: Query your document
docs = vectorstore.similarity_search("What is the Q3 revenue?", k=3)
for doc in docs:
    print(doc.page_content)

MCP Server Mode (LLM Agent Tool)

# Start PaddleOCR as an MCP (Model Context Protocol) server
$ paddleocr serve --port 8080 --lang en

# MCP server exposes OCR as a callable tool for LLM agents
# Compatible with Claude, ChatGPT, and any MCP-enabled agent
# Available tools: ocr_image, parse_pdf, extract_table

Python API Reference

PaddleOCR Class

PaddleOCR(lang, use_gpu, use_angle_cls, ...)

Main OCR pipeline class. Initializes text detection, angle classification, and text recognition models. Initialize once and reuse — model loading takes 1–3 seconds.

.predict(input)str | PIL.Image | np.ndarray

Run OCR on an image file path, PIL Image, or NumPy array. Returns a list of page results, each containing rec_texts, rec_boxes, and rec_scores.

PPStructureV3 Class

PPStructureV3(lang, use_gpu, ...)

Document layout analysis and structure pipeline. Handles PDFs (multi-page), images, and office documents.

.predict(input)

Parse a document. Returns list of page dicts with keys: markdown, blocks, page_id. Each block has type, bbox, and type-specific data.

Command Line Interface (CLI) Reference

PaddleOCR can be invoked from the command line using the paddleocr command after installation.

# Basic OCR on a directory of images
$ paddleocr --image_dir=./images --lang=en --use_gpu=False

# Structure analysis (detect tables, titles, layout)
$ paddleocr --image_dir=./docs --type=structure --lang=en

# Single image OCR
$ paddleocr --image_dir=image.jpg --lang=ch

# With angle classification (for rotated text)
$ paddleocr --image_dir=./scans --lang=en --use_angle_cls=True

# Start OCR HTTP server
$ paddleocr serve --port 8080 --lang en

CLI Flags Reference

Supported Language Codes

PaddleOCR supports 100+ languages. PP-OCRv6's unified model covers 50 languages out-of-the-box without any additional downloads. The following table lists commonly used language codes:

Troubleshooting Guide

This section covers the most common PaddleOCR issues and their solutions.

ImportError: No module named 'paddleocr'

Model download fails / timeout

CUDA out of memory

Poor accuracy on my documents

Changelog

v3.7.0 — June 11, 2026 (Latest)

• Released PP-OCRv6 with 3 model tiers (tiny 1.5M / small 7.7M / medium 34.5M)
• +4.6% detection, +5.1% recognition accuracy vs PP-OCRv5_server
• 50 language unified model (Chinese, English, Japanese, 46 Latin-script)
• 5.2× CPU speedup via OpenVINO; 6.1× on Apple M4 (tiny model)
• Improvements in digital displays, dot-matrix, tire prints, industrial text

v3.6.0 — May 28, 2026

• Released PaddleOCR-VL-1.6 — 96.33% on OmniDocBench v1.6 (SOTA)
• Official Python, Go, and TypeScript API SDKs released
• Multi-page TIFF file parsing support added

v3.5.0 — April 21, 2026

• Flexible inference backends: Paddle static, dynamic, and Transformers
• Word, Excel, PowerPoint → Markdown conversion
• DOCX export for parsed results
• PaddleOCR.js — official browser inference SDK for PP-OCRv5

v3.4.0 — January 29, 2026

• HPD-Parsing model added — 4,752 tokens/sec peak throughput
• MCP server mode for direct LLM agent tool use
• PP-ChatOCRv4 with improved document Q&A capabilities

View full changelog on GitHub →