RAG Testing
Also known as: Retrieval-Augmented Generation Testing
RAG testing verifies Retrieval-Augmented Generation systems — which fetch relevant documents before generating a response — checking both that retrieval finds the right source material and that the generated answer is actually grounded in what was retrieved, rather than drifting back to the model's own unverified internal knowledge.
RAG testing has two genuinely distinct failure points worth testing separately: retrieval failure (the system fetches the wrong or irrelevant documents for a given query) and generation failure (even with the right documents retrieved, the model still produces an answer that isn't actually faithful to them) — a system can fail at either stage independently.
Faithfulness testing specifically checks whether every claim in a generated answer can be traced back to the retrieved source material — a RAG system that occasionally ignores its retrieved context and answers from its own general training data instead is a subtle, easy-to-miss form of hallucination, since the answer might still be factually correct just not actually grounded in what was provided.
