Documents into structured Markdown for LLMs
docling solves the unglamorous but critical step before any RAG system: turning messy documents into clean, structured text a model can use. It takes PDF, DOCX, PPTX, HTML, images, and more, and produces Markdown or JSON that preserves reading order, headings, and tables rather than dumping a flat wall of text. It started at IBM Research (the DS4SD group) and is now hosted under the LF AI & Data Foundation as the docling-project, which is a meaningful governance signal for something you might build a pipeline on. At 61,641 stars as of 2026-06, it became one of the fastest-rising document tools of the RAG era.
What it does
The core job is layout-aware parsing. docling does not just extract characters; it understands page structure, so a two-column paper, a table, or a slide deck comes out with its structure intact. It can run OCR on scanned pages through pluggable backends, optionally use a vision-language model (GraniteDocling) for harder layouts, and it integrates with LangChain, LlamaIndex, and others, plus an MCP server for agents. Local execution is a deliberate feature, so sensitive documents never leave your machine.
Install
pip install doclingThen the CLI converts a file or URL straight to Markdown:
docling https://arxiv.org/pdf/2408.09869For harder documents you can switch on the vision-language pipeline, for example docling --pipeline vlm <file>. The detailed install options live in the project docs.
Where it fits, and where it doesn’t
Reach for docling when you need layout-aware parsing for a RAG pipeline, when tables and reading order matter, and when local, offline processing is a requirement for privacy. The foundation governance and MIT license make it a reasonable long-term dependency, and the LangChain and LlamaIndex integrations mean it drops into existing stacks.
It is heavier than a simple converter: it pulls in deep-learning models, so cold starts download weights and large PDFs cost real time and memory. If you only need quick text out of office files with no layout intelligence, a lightweight converter like markitdown is less work. And as the gotchas below show, table and scanned-document accuracy is good but not solved.
docling versus the other document parsers
| docling | unstructured | markitdown | jina reader | |
|---|---|---|---|---|
| Stars | 61,641 | 14,936 | 152,764 | 11,235 |
| Focus | layout-aware parsing | RAG preprocessing pipeline | lightweight conversion | URL to Markdown |
| Layout and tables | strong | partition-based | basic | n/a |
| Runs | local | local plus paid API | local | hosted plus self-host |
| Best for | papers, reports, scans | enterprise ETL | quick office to Markdown | web pages |
Counts are from GitHub as of June 2026. unstructured is the ETL-oriented pipeline with an open library and a paid platform. markitdown is the lightweight quick-conversion option. jina reader handles web URLs rather than local files. docling’s edge is layout-aware, local, foundation-governed parsing.
The star curve
The star curve shows a sharp rise as RAG went mainstream and “how do I get clean text out of a PDF” became everyone’s problem at once. The move under the LF AI & Data Foundation and the steady release cadence suggest the growth is being matched by real governance rather than a one-time spike.
What the issue tracker warns you about
The 900-plus open issues are a mix of feature requests and real limits worth knowing. Scanned and Chinese-language PDFs can still produce garbled output depending on the OCR backend, so test your specific documents. Table extraction is good but has had bugs where content is duplicated across cells, so verify tables that feed downstream logic. The dependency footprint is heavy and has drawn complaints about forced installs like OpenCV. And large documents are slow and memory-hungry, especially with the vision-language pipeline on. None of these undercut the core value; they are the cost of doing layout-aware parsing properly.
Related
For an ETL-oriented pipeline, see unstructured. For lightweight conversion, see markitdown. For web pages, see jina reader. For the OCR engines that back document parsing, see Tesseract and PaddleOCR. For what is trending, see the daily digest and the weekly report.
FAQ
docling vs unstructured, which should I use? docling focuses on layout-aware parsing that preserves reading order and tables, and runs locally. unstructured is an ETL-style pipeline with an open library plus a paid platform and broad RAG integrations. Choose docling for layout fidelity and local processing, unstructured for a managed ingest pipeline.
docling vs markitdown? docling does layout-aware parsing with OCR and optional vision models, at a heavier cost. markitdown is a lightweight converter for getting office files into Markdown quickly with no layout intelligence. Use docling when structure matters, markitdown when you just need text fast.
Can docling run offline? Yes. Local execution is a core feature, which is a key reason to choose it for sensitive documents. Note that first runs download models, so allow for that once.
Is docling accurate on tables and scanned PDFs? Good but not perfect. Tables are extracted with structure but have had duplication bugs, and scanned or Chinese PDFs depend on the OCR backend and can garble. Test your real documents before trusting the output downstream.
Who maintains docling? It started at IBM Research and is now hosted under the LF AI & Data Foundation as the docling-project, under an MIT license, which is a stronger governance signal than a single-company project.