Docling Documentation
by LF AI & Data / IBM Research
Turn PDFs, Office files and scans into clean, chunkable structure for RAG — entirely locally.
Overview
Docling is a Python library that turns messy real-world documents into structured, LLM-ready data, and this is its official documentation. It began in the AI for Knowledge team at IBM Research Zurich and is now hosted as a project of the LF AI & Data Foundation under an MIT licence; the repository has grown past 65,000 stars and requires Python 3.10 or newer. The docs are organised into Getting Started (installation, quickstart, GPU notes), Usage (advanced conversion options, supported formats, audio and video, enrichment features, vision models, the model catalog, an MCP server, agent skills, Jobkit and an API server), Concepts (architecture, the DoclingDocument model, serialization, confidence scores, chunking, plugins and OCR engines), Examples, Integrations and an API reference. Input coverage is unusually broad: PDF, DOCX, PPTX, XLSX, HTML, EPUB, LaTeX, Markdown, ODF, XBRL, email formats, images, audio and video. The headline capability is PDF understanding — page layout analysis, reading-order reconstruction, table structure recognition, code and formula detection, image classification and chart understanding — captured in a unified DoclingDocument and exported to Markdown, HTML, DocTags, WebVTT or lossless JSON. Visual Language Model pipelines, including IBM's GraniteDocling, and multiple OCR engines handle scanned material, and everything can run locally, which matters for air-gapped or regulated data. Plug-and-play integrations exist for LangChain, LlamaIndex, Haystack, Crew AI, Langflow, spaCy and NVIDIA. The underlying design is written up in the Docling Technical Report, arXiv:2408.09869.
At a Glance
- Topic
- RAG
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~2-3 hours for getting started plus the Concepts section; reference thereafter, self-paced
- Provider
- LF AI & Data / IBM Research
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Convert PDF, DOCX, PPTX, XLSX, HTML and EPUB files into one unified DoclingDocument
- ✓Extract table structure and reading order from complex PDFs instead of raw text dumps
- ✓Choose and configure OCR engines for scanned documents and image-only pages
- ✓Run Visual Language Model pipelines such as GraniteDocling on difficult layouts
- ✓Chunk parsed documents sensibly for embedding and retrieval-augmented generation
- ✓Export losslessly to Markdown, HTML, DocTags, WebVTT or JSON for downstream pipelines
- ✓Wire Docling into LangChain, LlamaIndex, Haystack or Crew AI RAG applications
- ✓Run the whole pipeline locally or air-gapped, and serve it via the MCP or API server
Highlights
- •Handles the formats other parsers skip — EPUB, XBRL, LaTeX, email, audio and video, not just PDF
- •DoclingDocument is a lossless intermediate representation, so you can re-export without re-parsing
- •MIT licensed with 65k+ GitHub stars and vendor-neutral governance under LF AI & Data
- •Local execution is a first-class documented path, which matters for regulated or air-gapped data
- •The design is backed by a citable technical report (arXiv:2408.09869), not just a README
Who It’s For
Best For
- ✓Engineers building RAG pipelines over PDFs and enterprise document stores
- ✓Teams that cannot send documents to a hosted parsing API for compliance reasons
- ✓Data engineers replacing brittle PyPDF or regex-based extraction code
Prerequisites
- •Python 3.10 or newer and comfort with pip or uv and virtual environments
- •Basic understanding of embeddings and retrieval-augmented generation to use the chunking docs
FAQ
What is Docling Documentation?
Official documentation for Docling, the document-parsing library that started at IBM Research Zurich and is now hosted by the LF AI and Data Foundation. It converts PDFs, Office files, HTML, images, audio and video into a structured DoclingDocument ready for retrieval-augmented generation, and covers layout and table extraction, OCR engines, vision-language pipelines, chunking for embeddings, fully local execution, and integrations with LangChain, LlamaIndex, Haystack and Crew AI.
Is Docling Documentation free?
Docling Documentation is free to access.
What level is Docling Documentation for?
Docling Documentation is aimed at a intermediate audience. Recommended background: Python 3.10 or newer and comfort with pip or uv and virtual environments, Basic understanding of embeddings and retrieval-augmented generation to use the chunking docs.
How long does Docling Documentation take?
Expect roughly ~2-3 hours for getting started plus the Concepts section; reference thereafter, self-paced. Most learners work through it at their own pace.
What will I learn from Docling Documentation?
You'll learn: Convert PDF, DOCX, PPTX, XLSX, HTML and EPUB files into one unified DoclingDocument; Extract table structure and reading order from complex PDFs instead of raw text dumps; Choose and configure OCR engines for scanned documents and image-only pages; Run Visual Language Model pipelines such as GraniteDocling on difficult layouts; Chunk parsed documents sensibly for embedding and retrieval-augmented generation; Export losslessly to Markdown, HTML, DocTags, WebVTT or JSON for downstream pipelines; Wire Docling into LangChain, LlamaIndex, Haystack or Crew AI RAG applications; Run the whole pipeline locally or air-gapped, and serve it via the MCP or API server.
Topics
Sources
This page was written from 4 sources, 3 on domains other than docling-project.github.io.