Week 18: Quiz (10 questions, 8/10 to pass)

1. MCQ: In the Foundry hierarchy, which resources does the hub own? (see Concepts §Model catalog / hub-project)

   • A) Deployments, evals, and traces
   • B) Azure OpenAI instances, AI Search indexes, storage, key vaults, and the network boundary
   • C) Only the agent definitions
   • D) Nothing; the project owns everything

2. MCQ: When you call the OpenAI SDK against Azure, the model= argument is your… (see notebook 01 cell 3)

   • A) Raw model ID, e.g. gpt-4.1
   • B) Deployment name you chose at deploy time
   • C) Region name
   • D) API version string

3. MCQ: Which deployment option bills whether you use it or not? (see Concepts §Deployment table)

   • A) Serverless endpoint (MaaS)
   • B) Provisioned throughput (PTU)
   • C) Pay-per-token on-demand
   • D) None of them

4. Short answer: In notebook 01, which two BoL fields does field_matches compare as integers (numeric comparison, not string equality)? (see notebook 01 cell 12)

5. MCQ: Which AI Gateway control is the right tool to cap monthly spend so a runaway agent cannot blow the budget? (see Concepts §AI Gateway)

   • A) Token-rate limit
   • B) Semantic cache
   • C) Token-usage quota / cost cap
   • D) Model routing

6. MCQ: In notebook 02's dry-run, which golden question fails before the fix, and why? (see notebook 02 cell 10)

   • A) Q1 (refund), because the prompt is wrong
   • B) Q4 (customs), because KEYWORD_DOCS has no customs → POL-004 mapping
   • C) Q2 (address), because $85 is misspelled
   • D) Q3 (dangerous goods), because UN number is missing from the doc

7. Short answer: Name at least five of the ten fields the Week 6 EXTRACT_PROMPT asks the model to return as JSON. (see notebook 01 cell 4)

8. MCQ: What is the recommended production authentication for Foundry code (instead of keys)? (see Concepts §How it breaks)

   • A) Hardcoded API key
   • B) Entra ID via DefaultAzureCredential / managed identity
   • C) A shared team password
   • D) No authentication at all

9. Short answer: Using the notebook's placeholder prices ($0.15/1M input, $0.60/1M output), compute the cost of one BoL extraction that uses 150 input and 120 output tokens. Show the arithmetic. (see notebook 01 cell 15)

10. MCQ: "A trace without a fix is a bug report" means… (see Concepts §Evaluation & tracing)

    • A) You should delete failed traces
    • B) Reading the trace is enough even if the score is unchanged
    • C) The eval loop is only complete when you read the trace, fix the worst failure, and show a before/after score
    • D) Traces are only for LLM-as-judge runs

Answer key

1. B: The hub owns shared enterprise resources; projects hold the team's deployments, evals, and traces. That split is the whole governance point.
2. B: In Azure, model= is your deployment name, not the raw model ID, the classic Week 18 gotcha.
3. B: PTU reserves capacity and bills regardless of usage; serverless is pay-per-token.
4. quantity and gross_weight_kg: both are compared with int(float(pred)) == int(float(truth)).
5. C: A token-usage quota / cost cap bounds cumulative spend; a rate limit only throttles the burst rate.
6. B: KEYWORD_DOCS deliberately omits customs/storage, so Q4 (customs hold fee) has no matching doc and fails until the fix adds customs → POL-004.
7. Any five of: shipper, consignee, port_of_loading, port_of_discharge, commodity, quantity, gross_weight_kg, declared_value_usd, freight_terms, date_of_issue.
8. B: Production guidance is managed identity / service principal via DefaultAzureCredential; keys are the legacy quick-start path.
9. Input 150 × $0.15/1M = $0.0000225; output 120 × $0.60/1M = $0.0000720; total ≈ $0.0000945 (under a tenth of a cent).
10. C: The loop closes only when the trace leads to a fix with a before/after number; a score without diagnosis, or a diagnosis without a fix, is incomplete.