Question 40
Domain 3: Application DevelopmentA RAG prompt includes retrieved context and asks the model to answer. Which instruction is most important for trustworthy behavior?
Correct answer: B
Explanation
Using only the retrieved context limits the model to grounded information, which reduces hallucinations. Saying when the context is insufficient is essential because a trustworthy RAG system should not invent facts beyond the provided evidence.
Why each option is right or wrong
A. "Always answer even when the context is missing."
Forcing an answer without context encourages fabrication instead of evidence-based responses.
B. "Use only the provided context when answering, and say when the context is insufficient."
In a retrieval-augmented generation setup, the model is being evaluated on whether it stays bounded by the supplied evidence rather than free-associating from parametric memory. The crucial instruction is to confine the answer to the retrieved passages and explicitly acknowledge insufficiency when the context does not support a conclusion; that is the core safeguard against unsupported claims and hallucinations in a RAG workflow.
C. "Prefer the longest possible answer."
Answer length does not improve factual grounding; verbosity can hide unsupported claims.
D. "Ignore source documents if the model seems confident."
Model confidence is not proof of correctness; source documents are the grounding signal.