Angela Shi and coauthor dissect the recurring failures in enterprise Retrieval‑Augmented Generation pipelines. They argue that most errors trace back to treating PDFs as plain text, discarding tables, layout, and hierarchy. The authors propose a four‑brick architecture that starts with a structured parser to preserve document semantics.

At scale, feeding whole documents to a LLM becomes financially untenable. A compliance team querying a 1,200‑page contract 200 times daily would spend roughly $131,000 annually using a naïve dump, versus a few hundred dollars with a scoped pipeline that retrieves only relevant pages. The cost gap widens dramatically across thousands of contracts.

The article warns against over‑tuning chunk size, overlap, or reranker thresholds while ignoring parsing quality. Teams waste weeks adjusting downstream knobs that cannot compensate for flattened tables or lost headings. Switching to a parser that outputs typed tables and page metadata restores precision and makes retrieval tuning meaningful.