Safe, Open, High-Performance — PDF for AI

OpenDataLoader-PDF converts PDFs into JSON, Markdown or Html — ready to feed into modern AI stacks (LLMs, vector search, and RAG).

It reconstructs document layout (headings, lists, tables, and reading order) so the content is easier to chunk, index, and query. Powered by fast, heuristic, rule-based inference, it runs entirely on your local machine and delivers high-throughput processing for large document sets. AI-safety is enabled by default and automatically filters likely prompt-injection content embedded in PDFs to reduce downstream risk.

• 🧾 Rich, Structured Output — JSON, Markdown or Html
• 🧩 Layout Reconstruction — Headings, Lists, Tables, Images, Reading Order
• ⚡ Fast & Lightweight — Rule-Based Heuristic, High-Throughput, No GPU
• 🔒 Local-First Privacy — Runs fully on your machine
• 🛡️ AI-Safety — Auto-Filters likely prompt-injection content - Learn more about AI-Safety
• 🖍️ Annotated PDF Visualization — See detected structures overlaid on the original

Download Annotated PDF Sample

• 🖨️ OCR for scanned PDFs — Extract data from image-only pages
• 🧠 Table AI option — Higher accuracy for tables with borderless or merged cells
• ⚡ Performance Benchmarks — Transparent evaluations with open datasets and metrics, reported regularly
• 🛡️ AI Red Teaming — Transparent adversarial benchmarks with datasets and metrics, reported regularly

• Java 11 or higher must be installed and available in your system's PATH.
• Python 3.9+

pip install -U opendataloader-pdf

• input_path can be either the path to a single document or the path to a folder.
• If you don’t specify an output_folder, the output data will be saved in the same directory as the input document.

import opendataloader_pdf opendataloader_pdf.run( input_path="path/to/document.pdf", output_folder="path/to/output", generate_markdown=True, generate_html=True, generate_annotated_pdf=True,
)

The main function to process PDFs.

Note: This package is a wrapper around a Java CLI and is intended for use in a Node.js backend environment. It cannot be used in a browser-based frontend.

• Java 11 or higher must be installed and available in your system's PATH.

npm install @opendataloader/pdf

• inputPath can be either the path to a single document or the path to a folder.
• If you don’t specify an outputFolder, the output data will be saved in the same directory as the input document.

import { run } from '@opendataloader/pdf'; async function main() { try { const output = await run('path/to/document.pdf', { outputFolder: 'path/to/output', generateMarkdown: true, generateHtml: true, generateAnnotatedPdf: true, debug: true, }); console.log('PDF processing complete.', output); } catch (error) { console.error('Error processing PDF:', error); }
} main();

run(inputPath: string, options?: RunOptions): Promise<string>

The main function to process PDFs.

Parameters

RunOptions

For various example templates, including Gradle and Maven, please refer to https://github.com/opendataloader-project/opendataloader-pdf/tree/main/examples/java.

To include OpenDataLoader PDF in your Maven project, add the dependency below to your pom.xml file.

Check for the latest version on Maven Central.

<project> <!-- other configurations... --> <dependencies> <dependency> <groupId>org.opendataloader</groupId> <artifactId>opendataloader-pdf-core</artifactId> <version>1.0.0</version> </dependency> </dependencies> <repositories> <repository> <snapshots> <enabled>true</enabled> </snapshots> <id>vera-dev</id> <name>Vera development</name> <url>https://artifactory.openpreservation.org/artifactory/vera-dev</url> </repository> </repositories> <pluginRepositories> <pluginRepository> <snapshots> <enabled>false</enabled> </snapshots> <id>vera-dev</id> <name>Vera development</name> <url>https://artifactory.openpreservation.org/artifactory/vera-dev</url> </pluginRepository> </pluginRepositories> <!-- other configurations... -->
</project>

To integrate Layout recognition API into Java code, one can follow the sample code below.

import org.opendataloader.pdf.api.Config;
import org.opendataloader.pdf.api.OpenDataLoaderPDF; import java.io.IOException; public class Sample { public static void main(String[] args) { Config config = new Config(); config.setOutputFolder("path/to/output"); config.setGeneratePDF(true); config.setGenerateMarkdown(true); config.setGenerateHtml(true); try { OpenDataLoaderPDF.processFile("path/to/document.pdf", config); } catch (Exception exception) { //exception during processing
        }
    }
}

The full API documentation is available at javadoc

Download sample PDF

curl -L -o 1901.03003.pdf https://arxiv.org/pdf/1901.03003

Run opendataloader-pdf in Docker container

docker run --rm -v "$PWD":/work ghcr.io/opendataloader-project/opendataloader-pdf-cli:latest /work/1901.03003.pdf --markdown --html --pdf

Build and install using Maven command:

mvn clean install -f java/pom.xml

If the build is successful, the resulting jar file will be created in the path below.

java/opendataloader-pdf-cli/target

java -jar opendataloader-pdf-cli-<VERSION>.jar [options] <INPUT FILE OR FOLDER>

This generates a JSON file with layout recognition results in the specified output folder. Additionally, annotated PDF with recognized structures, Markdown and Html are generated if options --pdf, --markdown and --html are specified.

By default all line breaks and hyphenation characters are removed, the Markdown does not include any images and does not use any HTML.

The option --keep-line-breaks to preserve the original line breaks text content in JSON and Markdown output. The option --content-safety-off disables one or more content safety filters. Accepts a comma-separated list of filter names. The option --markdown-with-html enables use of HTML in Markdown, which may improve Markdown preview in processors that support HTML tags. The option --markdown-with-images enables inclusion of image references into the output Markdown. The option --replace-invalid-chars replaces invalid or unrecognized characters (e.g., �, \u0000) with the specified character. The images are extracted from PDF as individual files and stored in a subfolder next to the Markdown output.

Options:
-o,--output-dir <arg>           Specifies the output directory for generated files
--keep-line-breaks              Preserves original line breaks in the extracted text
--content-safety-off <arg>      Disables one or more content safety filters. Accepts a comma-separated list of filter names. Arguments: all, hidden-text, off-page, tiny, hidden-ocg
--markdown-with-html            Sets the data extraction output format to Markdown with rendering complex elements like tables as HTML for better structure
--markdown-with-images          Sets the data extraction output format to Markdown with extracting images from the PDF and includes them as links
--markdown                      Sets the data extraction output format to Markdown
--html                          Sets the data extraction output format to HTML
-p,--password <arg>             Specifies the password for an encrypted PDF
--pdf                           Generates a new PDF file where the extracted layout data is visualized as annotations
--replace-invalid-chars <arg>   Replaces invalid or unrecognized characters (e.g., �, \u0000) with the specified character

Root json node

Common fields of content json nodes

Specific fields of text content json nodes (caption, heading, paragraph)

Specific fields of table json nodes

Specific fields of table row json nodes

Specific fields of table cell json nodes

Specific fields of heading json nodes

Specific fields of list json nodes

Specific fields of list item json nodes

Specific fields of header and footer json nodes

Specific fields of text block json nodes

We believe that great software is built together.

Your contributions are vital to the success of this project.

Please read CONTRIBUTING.md for details on how to contribute.

Have questions or need a little help? We're here for you!🤗

We love our brand and want to protect it!

This project may contain trademarks, logos, or brand names for our products and services.

To ensure everyone is on the same page, please remember these simple rules:

• Authorized Use: You're welcome to use our logos and trademarks, but you must follow our official brand guidelines.
• No Confusion: When you use our trademarks in a modified version of this project, it should never cause confusion or imply that Hancom officially sponsors or endorses your version.
• Third-Party Brands: Any use of trademarks or logos from other companies must follow that company’s specific policies.

This project is licensed under the Mozilla Public License 2.0.

For the full license text, see LICENSE.

For information on third-party libraries and components, see:

• THIRD_PARTY_LICENSES
• THIRD_PARTY_NOTICES
• licenses/