RAG quality is a chain, not a single score. A poor answer can originate in retrieval, chunking, ranking, context construction or generation.
Test retrieval separately
Start by checking whether the relevant source material is retrieved for representative questions. Include paraphrases, ambiguous queries and questions whose answer is intentionally absent.
Test grounding
Determine whether the generated answer is actually supported by the retrieved evidence. Unsupported claims should be identified even when the answer sounds confident.
Test faithfulness and completeness
A grounded answer can still omit important information or distort the source. Compare claims against evidence and check whether the response covers the required business facts.
Test missing knowledge
A robust RAG system should know when its knowledge is insufficient. Test whether it asks for clarification, declines, escalates or clearly states uncertainty rather than inventing an answer.
Need this applied to your AI system?
Turn the principle into an assessment, test strategy or engineering engagement.