Yellow.ai's D-RAC Chunker Ingests 795 PDF Pages in 72 Minutes, Cuts Chunking Tokens 95.7% vs Agentic Baselines

Yellow.ai's D-RAC paper (trending on Hugging Face) extends retrieval-aware chunking to arbitrary enterprise docs by normalizing everything to PDF, then doing a single Gemma-3 multimodal pass that emits retrieval-optimized Markdown with tables rewritten as per-row prose. On a 236-doc, 795-page benchmark it ingests the corpus in 72 minutes with zero errors, matching agentic chunking on retrieval quality while cutting chunking output tokens 95.7% and total chunking cost 77.8-85.6% under GPT-4.1 and Gemini 2.5 Pro pricing.