Techniques to Improve Generative AI Model Output →mediumMultiple Select

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

A data science team is building a document question-answering assistant on Vertex AI. Users report that answers are sometimes fabricated when the retrieved passages do not contain the answer. Which TWO techniques should the team apply to reduce hallucinations? (Choose two.)

⚠ Common exam trap

The trap here is believing that more randomness or a larger output budget improves reasoning, when both actually increase the chance of unsupported claims.

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

✓

Instruct the model to answer only from the provided context and to respond with a fixed phrase such as 'not found in the provided documents' when the context is insufficient.

Grounding instructions with an abstention path and mandatory citations both push the model to rely on retrieved passages and to expose when evidence is missing. Higher temperature and larger output limits increase unsupported text, while fine-tuning stores a stale snapshot without providing query-time provenance. Together, the two prompt and output controls reduce fabrication.

Answer analysis

Option-by-option breakdown

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

✗
Raise the maximum output token limit so the model has more space to explain its reasoning and avoid mistakes.
Why it's wrong here
A longer output limit only permits more text; it does not make the model more faithful to the retrieved context. Extra room can even invite more unsupported elaboration. Hallucination is a grounding problem, not a truncation problem, so increasing the token ceiling does not address the root cause.
✓
Instruct the model to answer only from the provided context and to respond with a fixed phrase such as 'not found in the provided documents' when the context is insufficient.
Why this is correct
An explicit grounding instruction with a defined abstention phrase gives the model a safe path when evidence is missing. It reduces fabrication because the model is told to prefer the supplied passages and to admit when they do not contain the answer. This is a low-cost prompt change that directly targets unsupported claims.
✗
Increase the temperature so the model explores alternative answers when the retrieved context seems incomplete.
Why it's wrong here
Higher temperature increases randomness and makes unsupported content more likely, not less. It encourages creative filling of gaps, which is exactly the behavior causing hallucinations. When the goal is faithfulness to retrieved passages, raising randomness works against grounding and should be avoided.
✓
Return citations to the specific retrieved passages used and require the answer to reference them.
Why this is correct
Citations force the model to tie each claim to a retrieved passage, which makes unsupported statements easier to detect and discourages invention. They also let users verify answers. By making provenance part of the required output, the model is nudged toward using only the evidence it was given.
✗
Fine-tune the model on the full document corpus so all answers are stored in the model weights.
Why it's wrong here
Fine-tuning embeds a snapshot and does not guarantee faithful retrieval at query time; it can still hallucinate and becomes stale as documents change. It also provides no citations. For a question-answering assistant over changing documents, grounding at inference time is more reliable than memorizing content in weights.

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