Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion

Document Retrieval-Aware Chunking (D-RAC). Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion. By Uday Allu and three others, from Yellow dot A I. Published on the twenty-first of September. No code repository is linked to it.

A retrieval-augmented generation system over an enterprise knowledge base has to ingest whatever documents the enterprise holds: PDFs, Word documents, presentations, scans. Their content is locked inside complex visual layouts, pages with several columns, and dense tables. The usual tools each fail in their own way. Rule-based extraction and OCR destroy the reading order, flatten the tables, and lose the hierarchy of headings. Fully agentic chunking over the extracted text avoids some of that, but costs a great many tokens and carries a risk of hallucination.

Document Retrieval-Aware Chunking, D-RAC, extends the authors' earlier Web Retrieval-Aware Chunking framework, W-RAC, from web content to documents of any format. Its first move exploits a simple fact: virtually every format has a faithful, deterministic rendering as a PDF. So D-RAC first normalizes any input document into a PDF, and from that point every format looks the same.

A single pass of a multimodal LLM then converts the rendered pages into Markdown optimized for retrieval. Tables are rewritten as self-contained statements in prose, and the heading hierarchy is preserved.

From there, chunking proceeds exactly as in W-RAC. The Markdown is parsed deterministically into units that can each be addressed by an ID. A lightweight LLM then plans the chunks, working over those identifiers rather than over the text. Source text is never regenerated during chunking. That keeps the benefits W-RAC already had, in cost, determinism and observability, while opening every renderable format as a first-class input.

The test set is the PDF subset of the RAG-Multi-Corpus benchmark: two hundred and thirty-six documents, seven hundred and ninety-five pages, spanning five enterprise domains. D-RAC converts and chunks the whole corpus in seventy-two minutes with zero errors, producing one thousand seven hundred and forty-eight chunks ready for retrieval.

Against agentic chunking with frontier LLMs, D-RAC cuts output tokens at the chunking stage by ninety-five point seven percent. Chunking cost falls by seventy-seven point eight percent at GPT four point one pricing, and by eighty-five point six percent at Gemini two point five Pro pricing. Chunking time drops by seventy-five percent. And D-RAC scales linearly to documents of more than five hundred pages.

The savings trace back to one rule: during chunking, source text is never regenerated. The document is converted to Markdown in a single multimodal pass, and after that the model planning the chunks handles only identifiers, which is why the output tokens shrink so sharply against agentic chunking, where the text itself passes through the model. Normalizing everything to a PDF first is the other half. It is what lets one pipeline, built for the web, take in every format that can be rendered.