Docling
Get your documents ready for gen AI
Last verified:
What is Docling?
Docling 2 is an open‑source document‑processing library that converts messy documents into structured, machine‑readable data so they can be more easily used by generative AI, RAG systems, and agentic workflows. It ingests a wide variety of formats—such as PDF, DOCX, XLSX, PPTX, HTML, LaTeX, Markdown, images, audio, and WebVTT—and normalizes them into a unified internal representation called the Docling Document. This representation preserves layout, reading order, tables, formulas, images, and auxiliary metadata, which makes downstream AI processing more accurate and consistent.
A key feature of Docling 2 is its advanced PDF understanding, including layout recovery, table structure, code blocks, mathematical formulas, and image classification. It also supports OCR for scanned pages and integrates with multiple Visual Language Models (such as GraniteDocling) to enhance visual content interpretation. The library can run locally or inside air‑gapped environments, which is useful when handling sensitive or regulated documents that cannot leave an internal network.
Docling 2 is targeted at developers, data engineers, and AI practitioners who need to ingest, clean, and structure documents before feeding them into language models, retrieval systems, or agents. It provides APIs and tooling for chunking documents into semantically coherent segments, exporting to formats like Markdown, HTML, plain text, WebVTT, and lossless JSON, and plugging into frameworks such as LangChain, LlamaIndex, Crew AI, and Haystack. Overall, it acts as a document‑preparation layer that sits between raw documents and generative‑AI pipelines.
Docling 2 is organized as a Python library with a CLI and complementary sub‑projects such as Docling Serve (for running Docling as an API service), Docling Core (which defines the Docling Document data model), and Docling Java (for JVM‑based environments). The project is hosted under the LF AI & Data Foundation and is maintained by the AI for knowledge team at IBM Research Zurich, which signals a focus on robust, research‑grade document understanding for enterprise and research use cases.
Docling pricing
Pricing model: Freemium
Docling 2 is open‑source and available under a permissive license, so the core library and most of its tooling are free to use, modify, and redistribute. There is no publicly listed paid tier or commercial license; instead the project is hosted under the LF AI & Data Foundation and offered as a community‑driven project. Commercial users may obtain additional support, integration help, or custom enhancements from the project maintainers or ecosystem partners, but the website does not specify fixed paid plans or subscription prices for Docling itself.
Docling pros
- Supports many input formats including PDF, DOCX, XLSX, PPTX
- Handles HTML, LaTeX, Markdown, and plain text files
- Supports images such as PNG, JPEG, TIFF, BMP, WEBP
- Supports audio formats like WAV, MP3, M4A, AAC, OGG, FLAC
- Supports video via audio‑track extraction and transcription
- Works with WebVTT for timed text and subtitles
- Supports domain‑specific XML schemas such as USPTO patents and JATS articles
- Supports XBRL for financial reporting documents
- Provides advanced PDF understanding with layout and reading order
- Recognizes table structures and formulas in PDFs
- Supports local execution for sensitive or air‑gapped data
- Offers extensive OCR for scanned PDFs and images
- Integrates with multiple Visual Language Models (for example GraniteDocling)
- Includes an audio‑processing module with ASR models
- Exposes a convenient CLI for quick document conversion
- Provides lossless JSON serialization of the Docling Document
- Offers exports to Markdown, HTML, Doctags, and plain text
- Provides plug‑and‑play integrations with LangChain and LlamaIndex
- Supports integration with agent frameworks such as Crew AI and Haystack
- Can be deployed as an API service via Docling Serve
- Offers a Java client for JVM‑based environments (Docling Java)
- Supports MCP‑based agent connectivity for agentic applications
- Includes structured information extraction (beta) for targeted field‑level parsing
- Uses a new layout model (Heron) for faster PDF parsing
Docling cons
- Currently in active development, so some features are marked as beta
- Structured information extraction is still in beta with evolving stability
- Chart and complex diagram understanding is listed as ‘coming soon’
- Chemistry‑specific structure understanding is not yet available
- Some advanced capabilities require additional dependencies or extra packages
- Running full OCR and Visual Language Model pipelines can be resource‑intensive
- Documentation emphasizes Python; non‑Python users rely more on the Java client or APIs
- Complex metadata extraction (title, authors, references, language) is still in progress
Frequently asked questions about Docling
What is Docling 2 and what does it do?
Docling 2 is an open‑source document‑processing library that converts documents from many formats into a unified, structured representation called the Docling Document. It normalizes PDFs, Office files, web pages, images, audio, video, and schema‑specific XML into a common model that preserves layout, reading order, tables, formulas, and other semantic elements, preparing documents for ingestion into generative‑AI and RAG pipelines.
Which document formats does Docling 2 support?
Docling 2 supports PDF, DOCX, XLSX, PPTX, HTML, XHTML, LaTeX, Markdown and its supersets, plain text, and various image formats such as PNG, JPEG, TIFF, BMP, and WEBP. It also handles audio files like WAV, MP3, M4A, AAC, OGG, FLAC, and video formats whose audio tracks it transcribes, plus WebVTT for timed text, and schema‑specific XML formats such as USPTO patents, JATS articles, and XBRL financial reports.
Can Docling 2 process scanned PDFs and images?
Yes, Docling 2 supports OCR for scanned PDFs and images, allowing it to recover text content from visually dense or non‑selectable documents. It can integrate with different OCR engines and Visual Language Models so that both layout and visual content are preserved in the resulting Docling Document.
What is the Docling Document data model?
The Docling Document is a unified, expressive data structure defined via pydantic that represents all document content and metadata in a single object. It includes fields for text blocks, tables, images, key‑value items, and layout‑related properties such as bounding boxes, while also encoding a hierarchical structure with body, furniture (headers and footers), and groups so that reading order and document organization are preserved.
What output formats can Docling 2 export to?
Docling 2 can export parsed documents into Markdown, HTML, plain text, WebVTT, Doctags (a layout‑aware markup format), and fully lossless JSON that serializes the entire Docling Document. These output options make it easy to feed the processed content into downstream tools like RAG systems, agents, or web‑based viewers.
Does Docling 2 support integration with AI frameworks?
Yes, Docling 2 provides plug‑and‑play integrations with popular AI frameworks such as LangChain, LlamaIndex, Crew AI, and Haystack. It also exposes an MCP server so that any agent can connect to Docling and leverage its document‑parsing capabilities inside agentic workflows.
Can I run Docling 2 locally or in air‑gapped environments?
Yes, Docling 2 is designed to run locally and can be deployed in air‑gapped environments where documents cannot leave a secure network. This makes it suitable for enterprises or organizations that need strong data‑privacy controls while still using advanced document‑parsing features.
Is there a command‑line interface for Docling 2?
Yes, Docling 2 includes a simple and convenient CLI that lets users convert documents directly from the terminal. For example, you can point the CLI to a URL, a local file path, or a directory, and it will process the documents and output them in one of the supported formats such as Markdown or JSON.
What are the upcoming features listed for Docling 2?
According to the project roadmap, upcoming features include metadata extraction such as title, authors, references, and language, enhanced chart understanding for bar charts, pie charts, line plots, and other visualizations, as well as deeper understanding of complex chemistry structures like molecular diagrams. These capabilities are listed as ‘coming soon’ and will further expand Docling 2’s role in scientific and technical document analysis.
Is Docling 2 free to use in commercial projects?
Docling 2 is open‑source and distributed under a permissive license, which allows free use in commercial projects without direct licensing fees. The project is hosted under the LF AI & Data Foundation and is maintained by the AI for knowledge team at IBM Research Zurich, focusing on community‑driven development rather than a closed‑source commercial product.