AI-900 Practice Question: Describe features of generative AI workloads on Azure
A company uses Azure OpenAI Service to power a chat-based support assistant. They have extensive knowledge base documents that contain the correct information. The company wants the assistant to answer questions solely based on the provided documents and avoid generating plausible-sounding but incorrect information. Which approach should they implement to minimize the risk of such fabrications?
⚠ Common exam trap
It's easy for candidates to assume prompt engineering (Option D) or fine-tuning (Option C) are sufficient to prevent hallucinations, but without retrieval-based grounding, the model can still generate confident-sounding falsehoods from its internal knowledge.
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) — provide relevant document excerpts as context in the prompt
Retrieval Augmented Generation (RAG) is the correct approach because it grounds the model's responses in actual, retrieved document excerpts provided as context in the prompt. This ensures the assistant answers based solely on the supplied knowledge base, directly minimizing the risk of hallucination (plausible-sounding but incorrect information) by constraining the model to the retrieved facts.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Retrieval Augmented Generation (RAG) — provide relevant document excerpts as context in the prompt
Why this is correct
Retrieval Augmented Generation (RAG) is correct because it injects relevant document excerpts directly into the model's prompt at query time, grounding the response in the retrieved content rather than relying solely on the model's parametric memory. This context-supplemented prompting lets the model answer based on the provided evidence, significantly reducing hallucinations and allowing the response to be traced to specific source passages without any model retraining.
- ✗
Increase the temperature parameter to 1.0 to force more creative responses
Why it's wrong here
Increasing the temperature parameter to 1.0 raises the randomness of token sampling, making the model more creative and disperse in its outputs — the opposite of what is needed for factually grounded answers. Higher temperature increases the probability of selecting lower-ranked, less likely tokens, which directly amplifies the chance of generating incorrect or irrelevant information; to reduce hallucinations, you would lower temperature toward 0 and provide relevant context.
When this WOULD be correct
In a scenario where the goal is to generate diverse, creative responses (e.g., a story generation tool) and factual accuracy is not required, increasing temperature to 1.0 would be appropriate to encourage varied outputs.
- ✗
Fine-tune the model on the knowledge base documents using supervised learning
Why it's wrong here
Fine-tuning the model on the knowledge base with supervised learning teaches the model patterns from those documents but does not create a hard guarantee that it will only answer from that learned material. The model still generates next tokens probabilistically, so when faced with a question that is only loosely related to the training data, it can produce plausible but fabricated facts — fine-tuning adjusts weights for style/behavior, not for verifiable retrieval at inference time.
When this WOULD be correct
When the goal is to adapt a base model to a specific domain or style (e.g., legal or medical terminology) using a fixed dataset, and the primary concern is improving accuracy on that domain rather than preventing hallucinations from external documents.
- ✗
Use prompt engineering with a system message that tells the model to never make up facts
Why it's wrong here
Using a system message that instructs the model to 'never make up facts' is a soft, heuristic constraint rather than a deterministic safeguard. Because the model has no internal fact-checking mechanism, it will still rely on its parametric knowledge and can fill gaps with confident-sounding fabrications when no external context is supplied; the instruction does not provide the grounding needed to suppress hallucination.
When this WOULD be correct
This option would be correct in a scenario where the model already has sufficient accurate knowledge (e.g., from pre-training or fine-tuning) and the goal is to enforce strict adherence to factual responses without relying on external data retrieval. For example, a question about reducing off-topic or speculative answers in a general-purpose chatbot where the model's training data is already reliable.
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.
✓Retrieval Augmented Generation (RAG) — provide relevant document excerpts as context in the promptCorrect answer▾
Why this is correct
Retrieval Augmented Generation (RAG) is correct because it injects relevant document excerpts directly into the model's prompt at query time, grounding the response in the retrieved content rather than relying solely on the model's parametric memory. This context-supplemented prompting lets the model answer based on the provided evidence, significantly reducing hallucinations and allowing the response to be traced to specific source passages without any model retraining.
✗Increase the temperature parameter to 1.0 to force more creative responsesWrong answer — click to see why▾
Why this is wrong here
Increasing the temperature to 1.0 makes the model more creative and random, which increases the likelihood of generating fabricated or incorrect information, contrary to the goal of minimizing fabrications.
★ When this WOULD be the correct answer
In a scenario where the goal is to generate diverse, creative responses (e.g., a story generation tool) and factual accuracy is not required, increasing temperature to 1.0 would be appropriate to encourage varied outputs.
Why candidates choose this
Candidates may think that higher temperature forces the model to rely less on its training data and more on provided context, but in reality it increases randomness and hallucination risk.
✗Fine-tune the model on the knowledge base documents using supervised learningWrong answer — click to see why▾
Why this is wrong here
Fine-tuning on the knowledge base documents does not inherently prevent the model from generating plausible-sounding but incorrect information; it can still hallucinate or produce fabrications, especially if the training data is incomplete or the model overgeneralizes.
★ When this WOULD be the correct answer
When the goal is to adapt a base model to a specific domain or style (e.g., legal or medical terminology) using a fixed dataset, and the primary concern is improving accuracy on that domain rather than preventing hallucinations from external documents.
Why candidates choose this
Candidates may believe that fine-tuning the model on the knowledge base will 'teach' it the correct information, thereby eliminating fabrications, but they overlook that fine-tuning does not guarantee grounded responses and can still lead to hallucinations.
✗Use prompt engineering with a system message that tells the model to never make up factsWrong answer — click to see why▾
Why this is wrong here
A system message telling the model to never make up facts is a form of prompt engineering, but it does not provide the model with actual correct information from the knowledge base. Without relevant context, the model may still hallucinate because it lacks authoritative source material to ground its responses.
★ When this WOULD be the correct answer
This option would be correct in a scenario where the model already has sufficient accurate knowledge (e.g., from pre-training or fine-tuning) and the goal is to enforce strict adherence to factual responses without relying on external data retrieval. For example, a question about reducing off-topic or speculative answers in a general-purpose chatbot where the model's training data is already reliable.
Why candidates choose this
Candidates may think that a strong instruction in the system message is enough to prevent hallucinations, underestimating the model's tendency to generate plausible-sounding but incorrect information when it lacks specific factual grounding.
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
Learn chapter
Azure Machine Learning Studio
Key term
Retrieval augmented generation
Retrieval augmented generation (RAG) is a technique that combines information retrieval with text generation to produce more accurate and contextually relevant answers.
Key term
RAG
Retrieval-Augmented Generation (RAG) is a technique that combines information retrieval with text generation to produce more accurate and contextually relevant answers from large language models.
About these practice questions
This AI-900 question is part of Courseiva's 985-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This AI-900 practice question is part of Courseiva's free Microsoft certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the AI-900 exam.