D-RAC Rebuilds the Entry Point for Enterprise RAG

Introduction

Enterprise knowledge bases rarely contain clean plain text. PDFs, DOCX files, slide decks, spreadsheets, and scans often combine multiple columns, page-spanning tables, headers, footers, and inconsistent heading levels. Conventional extraction can scramble reading order, flatten tables, and discard hierarchy before retrieval even begins. Document Retrieval-Aware Chunking, or D-RAC, addresses this ingestion problem rather than treating chunking as an isolated text-processing step.

How the pipeline works

D-RAC extends the earlier W-RAC approach from web content to arbitrary document formats that can be rendered. Its pipeline has three main stages:

• Format normalization: Every input is deterministically normalized into PDF and rendered as page images. The goal is to preserve the document’s visual organization instead of relying only on character extraction.
• Multimodal conversion: A single multimodal language-model pass converts the pages into retrieval-optimized Markdown. Heading levels are made explicit, decorative imagery is removed, and table rows are rewritten as self-contained prose that carries the relevant column headers.
• Identifier-based planning: The Markdown is parsed deterministically into addressable elements. A lightweight language model then plans chunks over element identifiers rather than reading and regenerating the full source text.

The table policy is particularly important. Disjunctive information such as “16 or 20 years” must not be merged into an ambiguous statement. Preserving values, conditions, and column context can determine whether a retrieved passage remains useful for downstream question answering.

Why the design matters

Many agentic chunking systems pay their largest cost when they regenerate an entire document. That stage consumes output tokens and creates opportunities for hallucinated or altered content. D-RAC limits multimodal interpretation to the conversion step, then keeps the later chunking stage source-preserving. This design is intended to improve cost control, determinism, and observability: each chunk can be traced back to structured elements rather than an opaque rewrite.

The supplied abstract reports a test on 236 documents and 795 PDF pages from the RAG-Multi-Corpus benchmark, spanning five enterprise domains. The full corpus subset was converted and chunked in 72 minutes with zero reported errors. However, the available material does not include the complete retrieval-quality comparison, so processing speed and pipeline structure alone are not enough to establish superiority over fixed-size chunking or conventional parsers.

The main risk moves upstream

D-RAC also makes the normalization stage a critical dependency. Tables that continue across pages, headings embedded in footers, merged cells, poor scans, and unusual layouts can still be misinterpreted. If the Markdown conversion damages the source structure, identifier-based planning may simply organize incorrect content more neatly. Production evaluation should therefore include weak-parser baselines, cross-page table cases, and end-to-end retrieval metrics—not only successful chunk creation.

The broader contribution is an engineering reordering of RAG ingestion: recover document structure with multimodal understanding first, then apply deterministic parsing and lightweight planning. That separation could make heterogeneous enterprise documents easier to place in a common retrieval pipeline, provided the conversion layer proves reliable on difficult layouts.

Source: Hugging Face Daily Papers