Docling

Docling — Document processing and understanding for gen-AI

Docling is an open-source toolkit and library that simplifies preparing documents for generative-AI applications. It focuses on robust parsing, structured representation, and integrations that let downstream systems (LLMs, retrieval systems, agents) consume documents reliably.

Key capabilities

• Parsing of many formats: PDF, DOCX, PPTX, XLSX, HTML, images (PNG, TIFF, JPEG, ...), audio (WAV, MP3), WebVTT, and more.
• Advanced PDF understanding: page layout & reading order, table structure recovery, code/formula detection, image classification and richer layout parsing.
• OCR support: processes scanned PDFs and images with OCR pipelines to extract text and layout.
• Audio/ASR support: transcription for audio inputs and support for WebVTT track parsing.
• Unified data model: a DoclingDocument representation that expresses structure, annotations and metadata in a consistent format and can be exported to Markdown, HTML, JSON, or DocTags.
• Local-first operation: supports local execution and air-gapped environments for sensitive data handling.
• MCP server: a lightweight server to connect Docling processing into agentic workflows and pipelines.
• Integrations: plug-and-play adapters for LangChain, LlamaIndex, Haystack and others to accelerate RAG and agent setups.

Highlights & design goals

• Developer-friendly CLI and Python API for quick conversion and experimentation.
• Extensible pipeline architecture to swap OCR engines, VLMs, or layout models.
• Focus on quality of document structure extraction (tables, code blocks, formulas), which improves downstream retrieval and LLM prompting.
• Works across major OSes (macOS, Linux, Windows) and CPU architectures (x86_64, arm64).

Typical usage

• Convert a PDF or URL to a DoclingDocument and export to Markdown/JSON for indexing into a retriever.

Example (Python):

from docling.document_converter import DocumentConverter
 
source = "https://arxiv.org/pdf/2408.09869"
converter = DocumentConverter()
result = converter.convert(source)
print(result.document.export_to_markdown())

CLI example:

docling https://arxiv.org/pdf/2206.01062

You can run with a visual-language model pipeline or enable OCR and other options via CLI flags.

Integrations & ecosystem

• Integrates with LangChain, LlamaIndex, Haystack and other agent/LLM frameworks for retrieval-augmented generation and agentic document use.
• Supports visual-language models (e.g., GraniteDocling) and can be combined with VLMs hosted on Hugging Face or local model runtimes.
• Provides output formats tailored for downstream pipelines (lossless JSON, Markdown, HTML, DocTags).

Origin, license & governance

• The project was started by the Deep Search Team / IBM Research Zurich and is hosted under the LF AI & Data Foundation.
• Licensed under MIT. Individual models used through the project retain their original licenses.

Technical report & docs

• A technical report documents inner workings and evaluation details (referenced in the project documentation).
• Full documentation, examples and integration guides are available on the project website and GitHub pages.

Roadmap & coming features

Planned improvements include richer metadata extraction (title, authors, references, language), chart understanding, and advanced domain-specific extractors (e.g., chemistry/molecular structures).

Who should use Docling

• Teams building RAG systems, document-centric agents, or search/indexing pipelines that need reliable structured document extraction.
• Organizations needing local/offline document processing for privacy-sensitive use cases.