Marker: High-Accuracy Document Conversion to Markdown and JSON

Introduction

Marker is a powerful, open-source tool developed by datalab-to, designed to convert various document types into structured formats such as Markdown, JSON, chunks, and HTML with high speed and accuracy. It supports a wide range of input files including PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB, across all languages.

Key features of Marker include:

• Conversion of diverse file types.
• Accurate formatting of tables, forms, equations, inline math, links, references, and code blocks.
• Extraction and saving of images.
• Intelligent removal of headers, footers, and other artifacts.
• Extensibility for custom formatting and logic.
• Structured extraction based on a JSON schema (beta).
• Optional integration with Large Language Models (LLMs) for boosted accuracy.
• Compatibility with GPU, CPU, and MPS for flexible deployment.

Marker has demonstrated favorable performance in benchmarks against leading cloud services like Llamaparse and Mathpix, as well as other open-source tools, offering superior speed and accuracy.

Installation

To get started with Marker, you'll need Python 3.10+ and PyTorch.

Install the core marker-pdf package:

pip install marker-pdf

For converting documents other than PDFs, install additional dependencies:

pip install marker-pdf[full]

Examples

Marker offers various ways to convert documents, from command-line interfaces to a Python API.

Interactive App

Try Marker interactively with a Streamlit app:

pip install streamlit streamlit-ace
marker_gui

Convert a Single File

Convert a PDF or image file from the command line:

marker_single /path/to/file.pdf

You can specify options like --output_format [markdown|json|html|chunks] and --use_llm for enhanced accuracy.

Convert Multiple Files

Process an entire folder of documents:

marker /path/to/input/folder

This command supports all the options available for marker_single.

Use from Python

Integrate Marker directly into your Python applications:

from marker.converters.pdf import PdfConverter
from marker.models import create_model_dict
from marker.output import text_from_rendered

converter = PdfConverter(
    artifact_dict=create_model_dict(),
)
rendered = converter("FILEPATH")
text, _, images = text_from_rendered(rendered)

Marker also provides specialized converters for specific tasks, such as TableConverter for extracting tables, OCRConverter for OCR-only processing, and ExtractionConverter for structured data extraction using a JSON schema.

Why Use Marker

Marker stands out as a robust solution for document conversion due to several compelling reasons:

• Unmatched Accuracy: Benchmarks show Marker consistently outperforms competitors in overall PDF conversion and table extraction, especially when augmented with LLMs.
• High Performance: It offers impressive throughput, processing documents rapidly on various hardware configurations, including GPUs.
• Versatile Input/Output: Supports a broad spectrum of document formats and provides flexible output options including Markdown, JSON, HTML, and optimized chunks for RAG applications.
• LLM Integration: The --use_llm flag allows leveraging powerful language models like Gemini, Google Vertex, Ollama, Claude, or OpenAI for superior accuracy in complex scenarios like table merging and form extraction.
• Extensible Architecture: Its modular design, based on Providers, Builders, Processors, and Renderers, makes it easy for developers to customize and extend its functionality.
• Commercial Options: For enterprise needs, Datalab offers a hosted API and an on-premise solution with high uptime and competitive pricing.

Links

• GitHub Repository: datalab-to/marker
• Datalab Platform: Datalab Platform
• Discord Community: Discord