docling

Docling is an open-source project initiated by the AI for knowledge team at IBM Research Zurich and is now hosted by the LF AI & Data Foundation. Its core goal is to simplify document processing by parsing diverse document formats into structured content suitable for use with generative AI.

Purpose: Docling is mainly used to parse documents and convert them into machine-readable formats for downstream use cases such as generative AI applications, retrieval-augmented generation (RAG), and agents. It places particular emphasis on advanced understanding of PDFs and supports multiple file types.

Main capabilities:

• Parse multiple document formats: including PDF, DOCX, PPTX, XLSX, HTML, EPUB, Apple Pages/Numbers/Keynote, WAV, MP3, WebVTT, Box Notes, email formats (EML, MSG), images (PNG, TIFF, JPEG, etc.), LaTeX, DocLang, plain text, and more.
• Advanced PDF understanding: including page layout, reading order, table structure, code, formulas, image classification, and more.
• A unified DoclingDocument representation format for easier downstream processing.
• Multiple export formats: Markdown, HTML, WebVTT, DocLang, DocTags, and lossless JSON.
• Support for specific XML schemas: DocLang, USPTO patents, JATS articles, XBRL financial reports.
• Local execution capability: suitable for sensitive data and air-gapped environments.
• Plug-and-play integration: integrates with agentic AI frameworks such as LangChain, LlamaIndex, Crew AI, Haystack, and more.
• Broad OCR support: for scanned PDFs and images.
• Support for multiple vision-language models (such as GraniteDocling).
• Audio support: using automatic speech recognition (ASR) models.
• Can connect to any agent via an MCP server.
• Can run as a service via an API server (docling-serve).
• Provides a simple and convenient CLI.
• Parse video files (MP4, AVI, MOV, MKV, WebM), extracting ASR transcripts and representative keyframes.
• Parse ODF files (.odt, .ods, .odp).
• Parse XBRL documentation.
• Parse email files (.eml, .msg).
• Parse EPUB files.
• Parse Apple Pages, Numbers, and Keynote documents.
• Parse plain text files (.txt, .text) and Markdown supersets (.qmd, .Rmd).
• Chart understanding: convert bar charts, pie charts, and line charts into tables or code, and add detailed descriptions.
• Metadata extraction: title, author, references, and language.
• Complex chemistry understanding: molecular structures.

Usage:

• Installation: requires Python 3.10 or higher (version 2.70.0 has dropped support for Python 3.9). Supports x86_64 and arm64 architectures on macOS, Linux, and Windows. Detailed installation instructions are in the documentation.
• CLI usage: convert documents via the command line and generate .md files containing structured content. Vision-language models such as GraniteDocling can also be used via the CLI.
• Python usage (recommended): provides more advanced usage and configuration options.
• Documentation and examples: the official documentation covers installation, usage, concepts, recipes, extensions, etc.; examples demonstrate different application scenarios.
• Integrations: native integration with popular frameworks and tools.

Applicable scenarios:

• Prepare document data for generative AI.
• Build RAG pipelines or agent applications.
• Process documents in sensitive data or air-gapped environments.
• Need to extract text and structure from scanned copies or images.
• Process complex PDFs containing tables, formulas, and code.
• Parse audio and video content to obtain transcriptions and key frames.
• Process domain-specific documents such as financial reports, patents, and academic articles.

Limitations:

• Requires Python 3.10 or higher.
• Some features depend on external models, and model licenses should be referenced from the original packages.
• Specific performance, accuracy, processing speed, etc. are not stated in the provided materials.

Other information: The Docling codebase is licensed under the MIT License. A technical report is available for reference. The project welcomes contributions and provides a discussion area for help and support.