A FAQ is already the answer, pre-written and paired with its question. When the source is already question-and-answer, the RAG has to treat it that way. This bonus article in the Enterprise Document Intelligence series by Angela Shi and Kezhan Shi explores how the standard four-brick RAG pipeline inverts when you design the corpus.

Parsing becomes trivial—loading a structured file with no PDF, layout reconstruction, or OCR needed. Retrieval doubles as a cache, and few-shot prompting becomes a retrieval problem. The authors analyze logs from a customer-support chatbot showing most queries are variations of the same fifteen questions, with answers already on disk.

The system pays generation costs unnecessarily when a lookup would suffice. Using a fifteen-entry synthetic FAQ for a fictional home-insurance product, the article demonstrates how architecture changes when the corpus is authored, not inherited. Runnable companion notebooks are available on GitHub at doc-intel/notebooks-vol1.

The team curating the FAQ chooses the schema, granularity, canonical phrasing, and tags—nothing must be recovered because nothing was lost.