Docling Documentation
by IBM Research / LF AI & Data Foundation
Turn messy PDFs, DOCX, slides and scans into clean, chunk-ready Markdown and JSON for your RAG pipeline.
Overview
Docling is an open-source document conversion toolkit created by IBM Research Zurich's AI for knowledge team and now hosted by the LF AI & Data Foundation under the MIT license. It parses PDFs, DOCX, PPTX, XLSX, HTML, EPUB, LaTeX, email formats (EML, MSG), images (PNG, TIFF, JPEG), audio (WAV, MP3) and specialized schemas such as XBRL financial reports and JATS articles, then exports them to Markdown, HTML, JSON or DocLang through a unified DoclingDocument representation. Its PDF pipeline does advanced layout analysis with reading-order reconstruction and table-structure recognition, plus OCR for scanned documents and vision-language-model parsing via GraniteDocling. Because it runs locally, sensitive documents never leave your infrastructure — a common blocker with hosted parsing APIs. The docs cover the Python API, the CLI, an API server, and first-party integrations with LangChain, LlamaIndex, Haystack and CrewAI, as well as MCP server support so agents can call document conversion as a tool. Recent additions include video file parsing (MP4, AVI, MOV, MKV, WebM), ODF support, chart understanding and email parsing.
At a Glance
- Topic
- RAG
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~2-4 hours to work through the guides and examples
- Provider
- IBM Research / LF AI & Data Foundation
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Convert PDFs, Office documents, HTML, images and audio into clean Markdown or JSON with the Docling Python API and CLI
- ✓Preserve document structure — reading order, headings, tables and figures — so retrieval chunks stay semantically coherent
- ✓Run OCR and vision-language (GraniteDocling) pipelines over scanned or image-heavy documents
- ✓Work with the unified DoclingDocument model and its chunking utilities to feed embeddings and vector stores
- ✓Wire Docling into LangChain, LlamaIndex, Haystack or CrewAI, and expose it to agents through its MCP server
- ✓Deploy conversion locally or behind the bundled API server to keep sensitive documents on your own infrastructure
Highlights
- •TableFormer-class table-structure recognition, not just naive text extraction — a common failure point in RAG ingestion
- •Fully local execution under an MIT license, so regulated and sensitive corpora never leave your environment
- •Hosted by the LF AI & Data Foundation with active 2026 development (video, ODF, chart understanding, email)
- •Drop-in integrations with the major RAG and agent frameworks plus MCP tool access
Who It’s For
Best For
- ✓AI engineers building RAG pipelines over PDFs, reports and enterprise document stores
- ✓Teams that cannot send documents to a hosted parsing API for privacy or compliance reasons
- ✓Data engineers preparing unstructured corpora for embedding and indexing
- ✓Agent developers who need document conversion as a callable tool
Prerequisites
- •Working Python knowledge
- •Basic familiarity with RAG concepts — chunking, embeddings, vector search
FAQ
What is Docling Documentation?
Official documentation for Docling, the open-source document-parsing toolkit that converts PDFs, Office files, HTML, images and audio into LLM-ready Markdown and JSON. It is aimed at AI engineers building RAG systems who need accurate layout, reading order and table extraction before chunking and embedding.
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: Working Python knowledge, Basic familiarity with RAG concepts — chunking, embeddings, vector search.
How long does Docling Documentation take?
Expect roughly ~2-4 hours to work through the guides and examples. Most learners work through it at their own pace.
What will I learn from Docling Documentation?
You'll learn: Convert PDFs, Office documents, HTML, images and audio into clean Markdown or JSON with the Docling Python API and CLI; Preserve document structure — reading order, headings, tables and figures — so retrieval chunks stay semantically coherent; Run OCR and vision-language (GraniteDocling) pipelines over scanned or image-heavy documents; Work with the unified DoclingDocument model and its chunking utilities to feed embeddings and vector stores; Wire Docling into LangChain, LlamaIndex, Haystack or CrewAI, and expose it to agents through its MCP server; Deploy conversion locally or behind the bundled API server to keep sensitive documents on your own infrastructure.