Enterprise Document Intelligence [Vol.1 #B00] reveals that retrieval-augmented generation (RAG) is not always the best solution. A support request lands in the queue, and the reflex is to prompt—but six cheaper methods often work better. An exact match handles clean identifiers, spelling fixes resolve simple errors, keyword search covers expert-vetted vocabulary, and embeddings address vocabulary gaps. Most requests are settled by these methods in milliseconds, each explaining its reasoning. The engineering challenge is reaching for the lowest rung that solves the case.

The post examines real document problems: classifying requests, matching free text to references, reading tables, cleaning OCR noise, and running models locally when data cannot leave. New to the series? Visit Angela Shi and Kezhan Shi's author pages on Towards Data Science for the full scope.

Six patterns emerged during work: spelling and OCR noise affect parsing, question parsing, and retrieval; 'I don't know' requests need evidence from all four bricks; tables require representation, retrieval, and generation choices; and faithful mocks matter across all contract-based components. A single article couldn't cover these adequately, so bonus articles provide the cross-cutting view.

Alternative pipeline shapes also deserve exploration. The FAQ corpus (team-designed) simplifies parsing and makes retrieval a cache. The dispatched architecture uses named routes over autonomous agents. Bonus articles walk these alternative shapes explicitly, not implicitly.