Docling

What is Docling?

Docling is an open source document-parsing toolkit, originally built by IBM Research, that converts PDFs, Office documents, HTML, and images into structured formats like Markdown, JSON, and HTML. It reads page layout, reading order, tables, code, and formulas, so the output stays usable in AI and retrieval pipelines. It’s a Linux Foundation AI & Data project released under the permissive MIT license.

What is Docling best for?

Teams building RAG systems, document search, or LLM pipelines who need clean, structured text out of messy real-world files — and want to run document parsing locally instead of shipping documents to a hosted API. It fits private or air-gapped environments especially well, since files never have to leave your network.

What can Docling do?

• Convert PDF, DOCX, PPTX, XLSX, HTML, EPUB, Markdown, and image files into Markdown, HTML, JSON, or DocTags
• Recognize page layout, reading order, and table structure, plus parse code blocks and formulas
• OCR scanned pages and images through pluggable OCR engines
• Understand charts and classify images, turning charts into tables or descriptions
• Run local vision-language models such as Granite-Docling for end-to-end parsing
• Plug into LangChain, LlamaIndex, Haystack, and Crew AI for document ingestion
• Serve it as a REST API with docling-serve, or connect agents through its MCP server

Where does Docling fall short?

• No forms or handwriting extraction. Docling targets printed and digital documents, so hand-filled forms and handwritten notes are weak spots.
• Best accuracy and the Granite-Docling vision model are compute-heavy. On a CPU-only machine, large or scanned PDFs parse slowly, and a GPU makes a real difference to throughput.
• It’s a toolkit you host yourself — there’s no official managed, SLA-backed cloud like LlamaParse, so deployment, scaling, and uptime are on you.

Is Docling free?

Yes — Docling is completely free and open source under the MIT license, with no paid tier, no per-page fees, and no usage caps. You can pip-install the Python package or run the containerized docling-serve API. Your only cost is the hardware it runs on, and a GPU helps with the heavier OCR and vision-model work.

What does Docling replace?

Docling is a self-hosted alternative to hosted document-parsing APIs like LlamaParse, which charge per page and require uploading your documents to their cloud. Docling does the parsing on hardware you control, with no per-page billing and no data leaving your network. It also overlaps with the open source Unstructured library for document ingestion, and it pairs naturally with Firecrawl, which does the equivalent job for web pages rather than files.

FAQ

Is Docling open source? Yes. Docling is released under the permissive MIT license and is hosted as a Linux Foundation AI & Data project, originally created by IBM Research. The source code and models are public.

Can I self-host Docling for free? Yes. Install the Python package or deploy the docling-serve container — both are free. You only pay for the server; CPU works, and a GPU accelerates OCR and the vision models.

Is Docling a good LlamaParse alternative? For local, privacy-sensitive parsing, yes — no per-page fees and documents stay on your own infrastructure. LlamaParse’s hosted API is simpler to start with and can edge ahead on some complex layouts.

What do I need to run Docling? Python 3.10+ for the library, or Docker/Podman for the docling-serve API. A GPU is optional but recommended for OCR-heavy or vision-model parsing at volume.