A new method tackles the enterprise RAG chunking problem while cutting cost by 95.7%.
Title: D-RAC: Document Retrieval-Aware Chunking
URL:
📌 Overview
D-RAC is a four-stage pipeline that normalizes messy enterprise documents (PDF, DOCX, PPTX, scans) into PDF, converts them once with a multimodal LLM into retrieval-optimized Markdown, then plans chunks deterministically over element IDs.
❗ The problem
Rule-based extraction destroys tables and heading hierarchy, while accurate agentic chunking regenerates the entire document, making it expensive at scale.
🛠️ Method
Tricks like turning each table row into a self-contained prose sentence, reconstructing heading hierarchy, and passing only element IDs to the planning LLM keep both retrieval quality and cost efficient.
📊 Results
Across 236 documents and 795 pages, D-RAC cut output tokens by 95.7% while Recall
@6 reached 0.798, beating agentic chunking (0.795) and rule-based extraction (0.717), with 75% less processing time.
🏢 Use cases
Performance stayed stable across automotive, banking, cloud, and other enterprise domains, making it a strong fit for production RAG pipelines.
#
RAG# #
DocumentAI#