If it searched my files, why can it still miss things?
Retrieval-augmented generation (RAG)
A system retrieves material and gives it to a model to help produce an answer. The retrieval and generation stages can each fail.
An everyday example
The assistant finds three policy extracts. Those extracts help answer a question, but do not prove that every policy was searched or that an exception was absent.
Original illustration, not an observed participant case.
Where the idea stops
Finding a document does not prove completeness, currency, permission to use it, or a correct interpretation.
Something to try
Ask what collection was searched, which passages were used, and what could be outside that scope. Verify a crucial passage yourself.
Practise on a fictional case →Go deeper: the builder’s view
For builders: evaluate retrieval coverage and answer faithfulness separately, with access controls applied before content reaches the model.
Source and interpretation
This is a plain-language interpretation with an applied example; it is not a quotation or an endorsement from the source.
Lewis et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) ↗Catalog v0.1 · checked 2026-10-29