Retrieval-augmented generation adds a retrieval step to an AI answering process. The system finds relevant content and uses it as context for generation. Microsoft describes RAG as grounding responses in an organisation’s content, while also identifying retrieval and access-control challenges. Source: Microsoft: retrieval-augmented generation

A travel example

Imagine an assistant answering a question about an agency’s opening hours using an approved page. In a useful test, the answer should match that page and identify it. Change the page, then test when the assistant starts using the updated information. This is a proposed evaluation, not a claim about any system’s refresh speed.

What to check

Ask which sources are included, who approves them and how outdated information is removed. Test questions with similar wording but different meanings so you can see whether the right passage is retrieved.

Grounding is not an accuracy guarantee

Evaluate the answer as well as the retrieved material. NIST identifies confident false output as a generative AI risk. A sensible acceptance test should therefore include missing information, conflicting sources and a question the assistant should pass to a person. Source: NIST: Generative AI Profile, AI 600-1

Written and updated by Travelgenix. Sources checked on 16 October 2026. Worked examples and checklists are editorial illustrations and recommendations, not customer results.