7.3. Mixed-Topic Practice Questions
These mirror the exam's scenario style. Read each, run the filters, then check the rationale.
Q1. A developer needs a support assistant to answer from monthly-updated product manuals, look up a customer's live order status, and respond in a fixed template. Which design is correct?
- A. Fine-tune a model on the manuals; add the template to the system prompt.
- B. A Foundry agent grounded on an indexed manual connection, a function tool for order lookup, with the template in instructions.
- C. A single chat completion with the manuals and template pasted into the prompt.
- D. A custom model trained on manuals plus a vision tool.
Answer: B. Manuals change monthly and are private → grounding/RAG, not fine-tuning (eliminates A and D — fine-tuning can't reliably recall facts or include new data). "Live order status" is an action against a system → a function tool, which requires an agent (eliminates C, a bare completion with no tool). The fixed template is durable behavior → instructions. B is the only option satisfying all three clauses.
Q2. A production app has steady, predictable traffic, must keep data processing within the EU, and must control its own encryption keys. Which combination?
- A. Standard / Global routing / Microsoft-managed keys.
- B. Provisioned (PTU) / Data Zone routing / customer-managed keys.
- C. Provisioned (PTU) / Global routing / customer-managed keys.
- D. Standard / Regional routing / Microsoft-managed keys.
Answer: B. "Steady, predictable" + production → Provisioned (eliminates A and D's Standard). "Within the EU" is a data-zone residency constraint → Data Zone routing (eliminates C's Global). "Control its own keys" → CMK (eliminates A and D's Microsoft-managed). Each requirement word maps to one axis.
Q3. A RAG assistant gives fluent, well-structured answers, but QA finds some claims aren't in the source documents. Which single control surfaces this?
- A. Coherence evaluator.
- B. Relevance evaluator.
- C. Groundedness evaluator.
- D. Raise the model temperature.
Answer: C. The answers are fluent (coherence is fine) and on-topic (relevance is fine) but unsupported by sources — that's precisely groundedness. A and B measure other dimensions and would pass these answers; D would make things worse. Groundedness is the RAG headline metric.
Q4. Users are crafting prompts that try to make the assistant ignore its safety rules. Which control addresses this?
- A. Lower the violence harm-category threshold.
- B. Prompt Shields.
- C. Groundedness detection.
- D. A TPM ceiling.
Answer: B. Jailbreak/injection attempts are exactly what Prompt Shields defend against. Harm-category thresholds (A) screen harmful content, not manipulation attempts; groundedness (C) catches hallucination; a TPM ceiling (D) caps spend. Only B matches the threat.
Q5. A team must strip Social Security numbers from millions of support transcripts before any generative model processes them, at low cost. Best approach?
- A. Prompt the LLM to ignore any PII it sees.
- B. Azure Language PII redaction as a preprocessing step.
- C. Fine-tune a model to avoid repeating PII.
- D. Store transcripts with customer-managed keys.
Answer: B. Privacy requires the model to never receive the PII, so redact first with the dedicated service — also far cheaper at millions-of-records scale than per-record LLM calls. A still sends PII to the model (privacy fail); C doesn't prevent intake; D encrypts at rest but doesn't redact content sent to the model.
Q6. An agent must compute statistics on an uploaded CSV and create a record in your CRM. Which tools?
- A. Code Interpreter for both.
- B. Function/connected tool for both.
- C. Code Interpreter for the statistics; a function/connected tool for the CRM record.
- D. File Search for the statistics; instructions for the CRM.
Answer: C. Code Interpreter runs sandboxed Python for analysis (statistics) but can't reach your CRM; a function/connected tool calls your CRM but isn't for ad-hoc computation. Each tool does its own job — the split is the point.
Q7. A pipeline must extract handwritten amounts from thousands of scanned forms with per-field confidence scores. Best fit?
- A. Prompt a multimodal model on each image.
- B. Azure Vision Read (OCR) model.
- C. A custom image-classification model.
- D. Image generation model.
Answer: B. Precise, high-volume printed/handwritten text extraction with confidence per field is exactly the Read/OCR model's purpose. Per-image multimodal prompting (A) is costly and lacks structured per-field confidence; classification (C) assigns categories, not text; generation (D) is the wrong direction entirely.
Q8. You want a generative model to classify tickets into Billing/Technical/Account, but it occasionally returns labels outside that set. The fix?
- A. Raise temperature for more options.
- B. Use structured output with an enum schema constraining the labels.
- C. Switch to a larger model.
- D. Add the categories to a fine-tuning dataset.
Answer: B. The output space is unconstrained, so the model invents labels — bound it with structured output (enum/schema) plus low temperature. A worsens it; C doesn't constrain output; D is heavyweight for what a schema solves instantly. (Remember: schema-valid still isn't fact-valid — validate.)
