AI-900 Practice Question: Describe features of generative AI workloads on Azure

Answer analysis

Option-by-option breakdown

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

• ✓

  Use Azure OpenAI On Your Data to connect to a product database so the model retrieves and references only the provided specifications.

  Why this is correct

  Azure OpenAI On Your Data integrates a retrieval pipeline—typically backed by Azure Cognitive Search—that indexes the product database and injects the most relevant specification chunks into the prompt context. The model then conditions its generation on those retrieved passages rather than relying on its parametric memory, so every generated claim is explicitly anchored to the provided specifications. This retrieval-augmented grounding is what directly mitigates hallucination for domain-specific product descriptions, because the output must trace back to the supplied source content.

• ✗

  Increase the frequency penalty to discourage the model from repeating common phrases.

  Why it's wrong here

  The frequency penalty adjusts logit scores by subtracting a value proportional to how often each token has already appeared, which suppresses repetition and promotes lexical diversity across the response. This parameter operates at the token-distribution level, affecting fluency and vocabulary variety, but it has no influence on whether the model's factual claims are grounded in provided data. Even with a high frequency penalty, the model can produce novel, varied, and confidently wrong specifications, because the penalty neither injects the product database nor validates statements against it.

  When this WOULD be correct

  A developer wants to reduce repetitive language in AI-generated marketing copy while still allowing creative variation. Increasing the frequency penalty would discourage the model from reusing common phrases, making outputs more diverse.

• ✗

  Decrease the temperature to 0 so the model always picks the most likely next token, making it more predictable.

  Why it's wrong here

  Setting temperature to 0 forces greedy token decoding, meaning the model always selects the highest-probability next token, which reduces randomness and makes outputs more deterministic. However, that probability distribution is still derived solely from the model's pre-training knowledge, not from the customer's product data, so the model can confidently pick the single most likely fabricated specification. Temperature controls only sampling stochasticity—it does not change the underlying knowledge source or constrain the model to external facts, so zero temperature cannot prevent hallucinated details.

• ✗

  Enable content filtering to block any outputs that contain harmful or biased language.

  Why it's wrong here

  Enabling content filtering applies Azure AI safety classifiers that detect and block outputs in categories like hate, violence, sexual content, and self-harm, protecting against harmful or biased language. These filters operate on toxicity and safety signals, not on factual accuracy, and they have no mechanism to compare a claim against a product database or verify that the description matches the given specifications. A response can be completely safe and still invent nonexistent features or wrong attributes, so content filtering is orthogonal to the problem of hallucination.

  When this WOULD be correct

  A developer needs to ensure that generated responses for a customer-facing chatbot do not contain offensive, biased, or inappropriate content, even if the model otherwise produces accurate information.

Option-by-option analysis

Why each answer is right or wrong

Understanding why wrong answers are wrong — and when they would be correct — is what separates a 750 score from a 900. The AI-900 exam frequently reuses these exact scenarios with slightly different constraints.

Analysis generated from the official AI-900blueprint and verified against question context. The “when correct” sections are what AI assistants cite when candidates ask “what’s the difference between these options?”

Go deeper

Related to this question