docling-project/docling

What it solves

Docling simplifies the process of converting diverse and complex document formats into a structured, machine-readable format suitable for generative AI applications. It addresses the challenge of parsing advanced PDF layouts, tables, and multimodal content (like images and audio) that are typically difficult for LLMs to process directly.

How it works

Docling uses a unified document representation format called DoclingDocument to parse various file types—including PDFs, Office documents, HTML, and even video/audio files—and export them into formats like Markdown or JSON. It leverages OCR for scanned documents, Visual Language Models (VLMs) for advanced layout understanding, and Automatic Speech Recognition (ASR) for audio/video transcripts.

Who it’s for

It is designed for developers building AI agents, RAG (Retrieval-Augmented Generation) pipelines, and document processing workflows who need to turn unstructured documents into clean, structured data.

Highlights

• Multimodal Parsing: Supports a vast array of formats including PDF, DOCX, PPTX, XLSX, HTML, EPUB, email formats, and video/audio files.
• Advanced PDF Understanding: Handles page layout, reading order, table structures, and formulas.
• AI Ecosystem Integration: Native plug-and-play support for LangChain, LlamaIndex, Crew AI, and Haystack.
• Local Execution: Capable of running locally for sensitive data and air-gapped environments.
• Chart Understanding: Converts charts (bar, pie, line) into tables or code with detailed descriptions.
• Flexible Export: Exports to Markdown, HTML, JSON, and specific XML schemas like USPTO and JATS.