RapidDoc - Fast Document Parsing System

July 12, 2026 · View on GitHub

Introduction

RapidDoc is a lightweight open-source framework focused on document parsing. It supports OCR, layout analysis, formula recognition, table recognition, reading-order recovery, and conversion from complex documents to Markdown, JSON, Word, and HTML.

RapidDoc also supports native parsing for docx/doc, pptx/ppt, and xlsx/xls files without using neural models.

The project is based on MinerU, with VLM components removed. It focuses on efficient pipeline-based document parsing and keeps good performance even on CPU.

Most default models come from the PaddleOCR PP-StructureV3 family, including OCR, layout analysis, formula recognition, reading-order recovery, and some table models. These models are converted to ONNX for efficient CPU/GPU inference.

RapidDoc also supports custom OCR, formula, and table models by implementing the CustomBaseModel.batch_predict method. Integration for the PaddleOCR-VL series is included.

Features

OCR recognition powered by RapidOCR, with multiple inference engines.
Layout analysis with PP-DocLayout models, including PP-DocLayoutV3 as the default.
Formula recognition with PP-FormulaNet_plus models.
Table recognition with table classification, wired-table recognition, and SLANET_plus/UNITABLE for wireless tables.
Reading-order recovery using layout model output or PP-StructureV3-style XY-cut logic.
Native Office parsing for Word, PowerPoint, and Excel files.
Markdown, Word, HTML, JSON, and image output workflows.
Docker and FastAPI deployment examples.

Installation

Install the CPU version:

pip install rapid-doc[cpu]

Install the GPU version:

pip install rapid-doc[gpu]

Install from source:

git clone https://github.com/RapidAI/RapidDoc.git
cd RapidDoc
pip install -e .[cpu]

For GPU development:

pip install -e .[gpu]

GPU Inference

Set the device before importing rapid_doc:

import os

os.environ["MINERU_DEVICE_MODE"] = "cuda"    # cuda:0
# os.environ["MINERU_DEVICE_MODE"] = "cuda:1"

The GPU package installs onnxruntime-gpu. Make sure your CUDA/cuDNN environment matches the ONNXRuntime CUDA execution provider requirements.

Quick Start

import os
from pathlib import Path

from rapid_doc import RapidDoc

root = Path(__file__).resolve().parent
output_dir = root / "output"

doc_path_list = [
    root / "demo/pdfs/示例1-论文模板.pdf",
    root / "demo/docx/test.docx",
]

engine = RapidDoc()
outputs = engine(doc_path_list, output_dir=output_dir)

for output in outputs:
    print(output.markdown)

RapidDoc - Fast Document Parsing System

Introduction

Features

Installation

GPU Inference

Quick Start

Docker

Online Demo

Examples and Parameters

Model Downloads

Benchmark

Acknowledgements

License