GCP Generative AI Leader Practice Question 38: RAG Output Retrieval Quality
Google Cloud Generative AI Leader · Test 4 · question 8 of 10
A product leader is deciding whether to add retrieval-augmented generation (RAG) to an internal chatbot that currently answers only from what the model learned during training. Which statement about how RAG changes the chatbot's generated output is accurate?
Show answer and explanation
Answer:
RAG first retrieves relevant facts for a question, then gives those facts to the model before it generates an answer. That means the response can reflect information published after the model's training cutoff and, with features such as high-fidelity grounding, can attach sources and confidence scores to the claims it makes. None of that removes the model's dependence on retrieval: if the retrieval step returns weak or irrelevant documents, the generated answer degrades along with it.
Option A overstates the benefit; retrieved documents can be outdated, irrelevant, or simply wrong, and grounding reduces but does not eliminate hallucination. Option C is wrong because RAG augments an LLM with retrieved context, it does not replace the LLM, which is still required to generate the final answer. Option D is wrong because RAG offerings such as prebuilt RAG with Agent Search and RAG APIs are built specifically to retrieve from unstructured content such as PDFs and web pages, not only structured tables.
Reference: RAG and grounding on Vertex AI