Choose the best answer for each question. Answers are at the bottom.
Q1. You need to analyze both an image and text in a single API call. Which model supports this?
A) GPT-4
B) GPT-4o (with vision)
C) DALL-E 3
D) text-embedding-ada-002
Q2. You need to process 50,000 customer feedback entries overnight. Which approach is most cost-effective?
A) Real-time endpoint with auto-scaling
B) Global Batch API
C) Multiple concurrent async calls
D) Azure Functions with HTTP triggers
Q3. How much cost savings does the Global Batch API offer compared to real-time endpoints?
A) 10%
B) 25%
C) 50%
D) 75%
Q4. Which technology enables progressive token display in chat applications?
A) WebSockets
B) Server-Sent Events (SSE) streaming
C) Long polling
D) gRPC streaming
Q5. Your chat app shows a blank screen for 8 seconds while waiting for a full response. How do you fix this?
A) Increase the timeout
B) Enable streaming to display tokens as they’re generated
C) Use a faster model
D) Cache all responses
Q6. What is the correct retry strategy when receiving a 429 (rate limit) response?
A) Retry immediately
B) Wait exactly 1 second and retry
C) Exponential backoff with jitter, or use the Retry-After header
D) Switch to a different model
Q7. Why is jitter important in retry logic?
A) It improves response quality
B) It prevents thundering herd — all clients retrying simultaneously
C) It reduces token usage
D) It is required by the Azure API
Q8. You want to track AI costs per department for chargeback. What is the recommended approach?
A) Create separate Azure subscriptions per department
B) Use APIM policies to tag requests and emit token metrics per department
C) Manually review invoices monthly
D) Use Azure Cost Management tags only
Q9. GPT-4o doesn’t accept video directly. How do you analyze video content?
A) Convert video to audio and transcribe only
B) Extract key frames as images and analyze each frame
C) Use Azure Video Analyzer instead
D) Upload the full video file to the API
Q10. Which Azure service provides global load balancing with WAF for a multi-region AI deployment?
A) Azure Load Balancer
B) Azure Traffic Manager
C) Azure Front Door
D) Application Gateway
Q11. Your application needs to handle a user input that exceeds the model’s context window. What should you do?
A) Increase the model’s context window
B) Truncate the input using token counting (e.g., tiktoken)
C) Send the request anyway and let the API handle it
D) Switch to a different model
Q12. What SLA does the Global Batch API guarantee for completion?
A) 1 hour
B) 4 hours
C) 24 hours
D) 48 hours
Q13. In a multi-region architecture, what happens when the primary Azure OpenAI region returns errors?
A) Requests queue until the region recovers
B) APIM’s circuit breaker fails over to the secondary region
C) Users see an error page
D) The application retries indefinitely
Q14. Which voice option is NOT available in Azure OpenAI’s TTS?
A) alloy
B) nova
C) cortana
D) shimmer
Q15. Your enterprise AI system must achieve 99.99% availability. Which architectural pattern is essential?
A) Single-region with auto-scaling
B) Multi-region active-active with Front Door and APIM
C) Single-region with premium SKU
D) Serverless functions in one region
| Q | Answer | Explanation |
|---|---|---|
| 1 | B | GPT-4o supports multi-modal input — images and text in a single request. |
| 2 | B | Global Batch API is designed for high-volume async processing at 50% lower cost. |
| 3 | C | The Batch API offers approximately 50% cost savings vs real-time endpoints. |
| 4 | B | SSE streaming delivers tokens progressively as they’re generated by the model. |
| 5 | B | Streaming shows tokens immediately as generated, eliminating the blank wait time. |
| 6 | C | Honor Retry-After headers first; otherwise use exponential backoff with jitter. |
| 7 | B | Jitter randomizes retry timing to prevent all clients from retrying simultaneously. |
| 8 | B | APIM token metrics with department dimensions enable automated cost allocation. |
| 9 | B | Extract frames at intervals and send as images — GPT-4o analyzes each frame. |
| 10 | C | Azure Front Door provides global load balancing, SSL termination, and WAF. |
| 11 | B | Use tiktoken to count tokens and truncate input to fit within the context window. |
| 12 | C | The Global Batch API guarantees completion within 24 hours. |
| 13 | B | APIM circuit breaker detects failures and routes traffic to the healthy region. |
| 14 | C | “cortana” is not an available TTS voice; the options are alloy, echo, fable, onyx, nova, shimmer. |
| 15 | B | Multi-region active-active with global load balancing is required for 99.99% SLA. |