OpenDataLoader PDF

OpenDataLoader PDF is an Apache-2.0 open-source PDF parser from the opendataloader-project organization. It converts PDFs into LLM-ready Markdown, JSON with bounding boxes, HTML, and Tagged PDF, and runs locally on CPU without a GPU. SDKs are available for Python, Node.js, and Java.

What It Is

OpenDataLoader PDF is a document-parsing library and CLI for developers building RAG and LLM pipelines. It takes digital, scanned, or tagged PDFs and produces structured output: headings, paragraphs, tables, lists, images, captions, and formulas, each with a page number and bounding box for source citations. It uses the XY-Cut++ algorithm for reading order in multi-column layouts and filters headers, footers, watermarks, and hidden text. It does not process Word, Excel, or PowerPoint files.

Local and Hybrid Modes

The default local mode is deterministic and rule-based. Hybrid mode runs a local AI backend, started with the opendataloader-pdf-hybrid command, that handles complex pages. It adds borderless-table handling, OCR for scanned PDFs (80+ languages, with selectable engines such as EasyOCR, RapidOCR, Tesseract, and Apple Vision), LaTeX formula extraction, AI descriptions of charts and images, and heading-hierarchy inference. The project publishes benchmark results in which its hybrid mode ranks first overall (0.907), a vendor-reported result.

Safety and Accessibility

Built-in AI safety filters remove hidden text, off-page content, and suspicious invisible layers that could carry prompt injection, and an optional sanitize flag replaces emails, URLs, and phone numbers with placeholders. For accessibility, an auto-tagging pipeline turns untagged PDFs into Tagged PDFs, following the PDF Association's Well-Tagged PDF specification and validated with veraPDF. The project says it was developed with Dual Lab. PDF/UA export and a visual accessibility editor are listed as enterprise add-ons available on request.

Setup Path

Install with pip install opendataloader-pdf (Java 11+ is required), call convert() on a batch of files, and pick output formats such as markdown, json, html, or tagged-pdf. A LangChain document loader is available as a separate package.