Most LLM training data and enterprise RAG pipelines depend on document ingestion, yet traditional tools waste tokens on lossy formats and struggle with real-world PDFs. The cost compounds at scale: inefficient document representation means higher training costs, larger context windows consumed, and degraded retrieval quality.
In this talk, I'll introduce Docling, an open-source document processing engine that takes a different approach. Docling uses purpose-built deep learning models to parse documents the way humans read them, preserving hierarchy, extracting tables, and maintaining reading order. I'll also introduce DocTags, a markup language we designed specifically for LLM tokenizers. DocTags uses 30-45% fewer tokens than HTML for the same content, and each tag maps directly to a single token, so you're not wasting context window space on markup overhead. We'll look at how IBM and Hugging Face used Docling to process 475 million PDFs for the FinePDFs dataset, extracting 3 trillion tokens at 50x lower cost than VLM-based approaches. I'll also cover IBM's work through LF AI & Data to establish DocTags as an ISO standard, creating a shared representation format that's optimized for both LLM training and inference.
Docling is fully open-source under MIT license, so you can run it locally for sensitive data and air-gapped environments.
Mingxuan Zhao is a Software Developer II at IBM working in Open Technology. A graduate of the Cockrell School of Engineering at the University of Texas at Austin, Mingxuan focuses on building developer-focused tools and contributing to modern web and cloud technologies. Based in New York, Mingxuan works on open and collaborative engineering initiatives within IBM.
