Techniques to Improve Generative AI Model Output →mediumMultiple Choice

Generative AI Leader Practice Question: Techniques to Improve Generative AI Model Output

A product team at a retailer is using Vertex AI Studio to build a Gemini-powered assistant that answers questions about their internal return policy. Early tests show the model invents policy details such as a 45-day return window, even though the official policy allows only 30 days. The team wants the assistant to answer strictly from a set of approved policy PDFs stored in a Cloud Storage bucket, and they want to avoid retraining the model. Which technique should they use?

⚠ Common exam trap

The trap here is assuming that lowering temperature or adding a firm instruction eliminates hallucination, when only supplying authoritative source content through grounding actually fixes a missing factual detail.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

✓

Retrieval-augmented generation (RAG) with a Vertex AI Search data store grounded on the approved policy documents.

Grounding with retrieval-augmented generation lets the assistant fetch the relevant passages from the approved policy PDFs and condition its answer on that retrieved text, so the 30-day window is stated correctly and the response can cite the source. It requires no retraining, updates automatically when documents change, and directly addresses hallucinated policy details.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✗
Adding a system instruction telling the model to be accurate and to never hallucinate policy details.
Why it's wrong here
A system instruction shapes tone and behavior but cannot supply the missing 30-day fact. Instructions alone do not give the model access to the approved PDFs, so it may still invent or guess policy numbers. This approach also provides no verifiable source citation, making it unsuitable when the team needs answers grounded in specific stored documents.
✗
Supervised fine-tuning of the Gemini model on a labeled dataset of correct policy answers.
Why it's wrong here
Supervised fine-tuning changes the model's weights to learn new patterns, which is heavier than needed here and requires curated labeled examples. More importantly, fine-tuning does not guarantee the model will quote exact policy numbers correctly and it does not automatically refresh when the policy PDFs change, so it fails the requirement to answer strictly from the approved documents without retraining.
✗
Lowering the model's temperature to 0 and increasing the top-k value in the generation configuration.
Why it's wrong here
Reducing temperature makes sampling more deterministic but does not inject factual policy content the model never learned. The model can still confidently produce the wrong 45-day window because decoding parameters only affect randomness, not knowledge. Increasing top-k widens sampling, which is the opposite of what is wanted, and neither setting grounds responses in the stored PDFs.
✓
Retrieval-augmented generation (RAG) with a Vertex AI Search data store grounded on the approved policy documents.
Why this is correct
RAG retrieves relevant passages from the indexed policy PDFs at query time and passes them to Gemini as grounding context, so answers reflect the approved 30-day policy instead of the model's pretrained assumptions. Vertex AI Search provides managed indexing and grounding, and no model retraining is required, which matches the team's constraint.

Go deeper

Generative AI Leader Practice Question: Techniques to Improve Generative AI Model Output

Why each option matters

Option-by-option breakdown

Related to this question