What's this about?
D-RAC renders varied office documents as PDFs, converts layouts and tables into retrieval-oriented Markdown, then plans chunks over identifiers.
What exactly is retrieval-aware chunking?
It splits content so each piece carries enough context to answer likely questions. A bad split can separate a table value from its heading and make correct retrieval impossible.
Why does this matter?
Author-reported preprint; 20 HF upvotes at fetch time.
- Enterprise files have layouts that plain OCR destroys.
- Source text is not regenerated during the chunking stage.
- Reported cost reductions need testing on noisier private corpora.