Enterprise retrieval-augmented generation (RAG) systems must ingest a wide range of document formats — PDFs, Word files, presentations, spreadsheets and scans — whose content is often locked inside complex visual layouts, multi-column pages and dense tables. According to a new paper posted on Hugging Face, rule-based extraction and OCR tend to destroy reading order, flatten tables and lose heading hierarchy, while fully agentic chunking over extracted text carries high token costs and hallucination risk.
To address this, the authors present Document Retrieval-Aware Chunking (D-RAC), an extension of their earlier Web Retrieval-Aware Chunking (W-RAC) framework to arbitrary document formats. D-RAC first normalizes any input document into PDF, relying on the observation that virtually every format has a faithful, deterministic PDF rendering. A single multimodal LLM pass then converts the rendered pages into retrieval-optimized Markdown, rewriting tables as self-contained prose statements and preserving heading hierarchy.
Chunking then proceeds exactly as in W-RAC: deterministic parsing into ID-addressable units, followed by lightweight LLM-based chunk planning over identifiers rather than raw text. Because source text is never regenerated during chunking, the paper says the approach preserves W-RAC's cost, determinism and observability benefits while making every renderable format a first-class input.
The authors report results on the 236-document, 795-page PDF subset of the RAG-Multi-Corpus benchmark, spanning five enterprise domains. D-RAC converted and chunked the entire corpus in 72 minutes with zero errors, producing 1,748 retrieval-ready chunks. Compared with agentic chunking using frontier LLMs, the method reduced chunking-stage output tokens by 95.7%, cut chunking cost by 77.8% under GPT-4.1 pricing and 85.6% under Gemini 2.5 Pro pricing, and reduced chunking time by 75%. The paper also states that D-RAC scales linearly to documents of 500+ pages.
In the paper's discussion thread, one commenter noted that the PDF parser ultimately determines what the retriever sees, and argued that retrieval quality should be measured against a bad-parser baseline, not only against naive chunking. The authors' framing emphasizes that the multimodal model touches content only once, with source text never regenerated.