Docling

What is Docling?

Docling is an open-source document processing project for generative AI, initiated by the AI for Knowledge team at IBM Research Zurich and currently part of the LF AI & Data Foundation project ecosystem. It addresses an often-overlooked step in knowledge base construction: PDFs and office documents contain more than just text—they also include fields, tables, images, and reading order—and simply copying text often causes these relationships to be lost.

Docling converts various types of input into a unified document structure, then exports it as Markdown, HTML, JSON, or searchable snippets. As of October 26, 2026, the official GitHub repository had accumulated approximately 69,600 stars. The project code is licensed under the MIT License, but the rights to the models, third-party dependencies, and source files used must be verified separately.

Key Features

• Multi-format parsing:It supports common file formats such as PDF, DOCX, PPTX, XLSX, HTML, and images, making it ideal for the centralized processing of documents from various sources within an organization.
• Layout and Reading Order:Focus on multi-column layouts, heading hierarchies, and element relationships to minimize the content confusion caused by extracting content line by line.
• Forms and Scanned Documents:Provides form structure recognition and OCR capabilities to deliver more usable input for reports, manuals, and scanned documents.
• Unified Document Object:Save the structure as a DoclingDocument so you can export it later as needed, without having to parse the original file each time.
• AI Workflow Integration:It can integrate with ecosystems such as LangChain, LlamaIndex, and Haystack, as well as projects like Docling Serve and MCP, making it easy to package translation capabilities into services.

Who is it for, and what value does it offer?

If a team is adding product manuals, corporate policies, and research papers to a knowledge base, Docling can handle the parsing process “before documents enter the search system.” It is particularly valuable in scenarios requiring table formatting and source attribution, as preserving the original structure makes it easier to trace back to the source during subsequent searches and responses. For the average reader, it is not a ready-to-use chat platform; rather, it is better suited as a data processing component for developers.

Documents such as contracts and bills can also be fed into the subsequent processing workflow; however, parsing is not the same as legal review, financial accounting, or factual determination. Significant amounts, clauses, and tables should continue to be checked manually or by specialized verification programs.

How to use

1. Set up Python 3.10 or later, install docling in a standalone environment, and start by using the official examples and non-sensitive samples.
2. Convert local files using the CLI, or use `DocumentConverter.convert` in Python, followed by calling export methods such as `export_to_markdown` on the document object.
3. Select the appropriate processing path—PDF, OCR, or visual language model—based on the input; there is no need to enable all extensions from the start.
4. Check the heading hierarchy, cross-page tables, reading order, and scanned text, then feed the output into vector search or other data pipelines.
5. Consider using Docling Serve when multiple users need to access the service; in production environments, implement task queues, resource limits, isolation, and logging—do not use the local example directly as a production configuration.

Comparison of Similar Projects

The selection process should be based on the input data and objectives: If you already have a large number of PDFs, prioritize evaluating document parsing solutions; if you need to collect web pages, consider scraping components first; and if you require a functional Q&A interface, you’ll also need an application layer. These three approaches are not entirely interchangeable.

Deployment, Costs, and Limitations

Local use does not require uploading files to a public service, but the model may need to be downloaded the first time the program is run; in offline or sensitive environments, model files should be prepared in advance, and you should confirm that remote models or external services are not enabled. The MIT license does not imply that all models are licensed under the same terms, nor does it exempt you from copyright and privacy responsibilities regarding input data. Fees may still apply for servers, computing power, and hosting services.

Low-quality scans, mixed-language layouts, and complex tables may affect results; 100% accuracy cannot be guaranteed. Extensions such as visual models and audio may require additional hardware and dependencies. Version v2.142.1, released on October 23, 2026, includes fixes for resource boundary and parsing robustness; when processing untrusted files, ensure the software is kept up to date and limit resource consumption.

Official Portal

GitHub Repositories - Official Documentation - official website - Release Notes - MIT License

Screenshots and logos from the official website: Docling official website and project repository; used solely for project introduction purposes.