Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion
D-RAC turns varied enterprise files into retrieval-ready chunks by first rendering them as PDFs, then converting pages into structured Markdown.
The paper says this avoids common ingestion failures around reading order, tables, and heading hierarchy. Chunking is then planned over document-unit identifiers rather than regenerated source text, keeping the process deterministic and cheaper than fully agentic chunking. On a 236-document PDF subset of RAG-Multi-Corpus, the authors report 72 minutes, zero errors, and 1,748 chunks. They also report a 95.7% reduction in chunking-stage output tokens versus frontier-LLM agentic chunking. HF Daily Papers' note
The paper says this avoids common ingestion failures around reading order, tables, and heading hierarchy. Chunking is then planned over document-unit identifiers rather than regenerated source text, keeping the process deterministic and cheaper than fully agentic chunking. On a 236-document PDF subset of RAG-Multi-Corpus, the authors report 72 minutes, zero errors, and 1,748 chunks. They also report a 95.7% reduction in chunking-stage output tokens versus frontier-LLM agentic chunking. HF Daily Papers' note
score 4