Docling parses documents in many formats into a unified DoclingDocument representation and exports them for generative-AI applications. Its PDF understanding recovers page layout, reading order, table structure, code, formulas, and image classes, and OCR handles scanned PDFs and images.
Docling runs fully locally, which suits sensitive data and air-gapped environments. You use it as a Python library, from a CLI, as an API server (docling-serve), or from any agent through its MCP server. Visual language models such as GraniteDocling can replace the default conversion pipeline.
Recent additions include video parsing with an ASR transcript and representative keyframes, OpenDocument (ODF), XBRL financial reports, Apple Pages, Numbers and Keynote files, and chart understanding that converts bar, pie, and line charts into tables or code with descriptions. Metadata extraction and chemistry (molecular structure) understanding are announced as coming next.
Features
- Input formats: PDF, DOCX, PPTX, XLSX, HTML, EPUB, Apple Pages, Numbers and Keynote, WAV, MP3, WebVTT, Box Notes, EML and MSG email, images, LaTeX, DocLang, plain text, and more
- Export formats: Markdown, HTML, WebVTT, DocLang, DocTags, and lossless JSON
- PDF understanding: page layout, reading order, table structure, code, formulas, and image classification
- XML schemas: DocLang, USPTO patents, JATS articles, and XBRL financial reports
- OCR and VLMs: OCR for scanned PDFs and images, plus visual language models such as GraniteDocling
- Audio and video: ASR models transcribe audio, and video files yield a transcript and representative keyframes
- Chart understanding: converts bar charts, pie charts, and line plots into tables or code with detailed descriptions
- Framework integrations: LangChain, LlamaIndex, CrewAI, and Haystack
- Agent and service access: an MCP server for agents and an API server (docling-serve) for running Docling as a service
- Local execution: runs entirely on your own machine for sensitive data and air-gapped setups
