Docling turns PDFs, scans, slides and other messy inputs into a single structured document format that can be exported as Markdown, JSON, HTML or typed schema outputs.
More than 10 input formats are supported, including PDF, DOCX, PPTX, HTML, images and audio, while the pipeline handles OCR, reading order, table reconstruction, headers, captions and bounding boxes.
A Python install and four-line conversion flow let users process documents locally by default, avoiding API keys and keeping sensitive files such as contracts or medical records on-device.
Schema-based extraction is the key step: Docling's DocumentExtractor can return validated Pydantic fields like invoice totals and bill numbers instead of raw text that still needs parsing.
The open-source project began at IBM Research Zurich, is now under LF AI & Data, has native RAG-framework integrations, and gained a managed watsonx SaaS offering in June 2026.