D-RAC normalizes every enterprise document format to PDF before a single multimodal LLM pass

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

D-RAC attacks RAG ingestion over heterogeneous enterprise formats by exploiting the fact that virtually every format (Word, PowerPoint, scans) has a faithful deterministic PDF rendering, so it converts everything to PDF first and then runs one multimodal LLM pass to produce retrieval-optimized Markdown. The pitch against the two existing options is specific: rule-based extraction and OCR destroy reading order, flatten tables and lose heading hierarchy, while fully agentic chunking over extracted text costs too many tokens and invites hallucination. Notably it rewrites tables as self-contained prose statements rather than preserving table structure, which is the design choice worth arguing with if you run a document RAG pipeline.