Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion
D-RAC converts enterprise documents into retrieval-ready Markdown chunks, cutting chunking tokens 95.7% versus agentic methods.
D-RAC extends Web Retrieval-Aware Chunking to heterogeneous enterprise files by rendering every input as PDF, then using one multimodal LLM pass to produce retrieval-oriented Markdown that preserves headings and rewrites tables as self-contained prose. Chunking then follows deterministic ID-addressable parsing and lightweight planning over identifiers, so source text is not regenerated. On a 236-document, 795-page PDF subset of RAG-Multi-Corpus it created 1,748 chunks in 72 minutes with zero errors. Compared with frontier agentic chunking, output tokens fall 95.7 percent, cost falls 77.8 to 85.6 percent, and time falls 75 percent, with linear scaling past 500 pages.
- Any renderable format is normalized to PDF, then converted in one multimodal pass.
- Chunk planning uses identifiers so source text is never regenerated.
- 236 documents and 795 pages yield 1,748 chunks in 72 minutes with zero errors.
- Chunking output tokens fall 95.7 percent versus frontier agentic chunking.
- Cost drops 77.8 percent for GPT-4.1 and 85.6 percent for Gemini 2.5 Pro.
- AI models
- GPT-4.1Gemini 2.5 Pro
From the original article
1 min read at huggingface.co
Retrieval-Augmented Generation (RAG) systems over enterprise knowledge bases must ingest heterogeneous document formats -- PDFs, Word documents, presentations, and scans -- whose content is locked inside complex visual layouts, multi-column pages, and dense tables.
By Uday Allu, Abhivanth Sivaprakash, Pratik Singh · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.