AI Intermediate · 22% of the exam

Azure AI Foundry & Azure OpenAI: free practice questions

5 sample questions from our 165-question bank for this domain — answers and explanations included. These are the same scenario-based style as the real DataCertPrep exam.

1. An AI engineer wants to deploy an open-source Llama 3 model from the Azure AI Foundry model catalog to a managed online endpoint. After selecting the model, which deployment target within Azure AI Foundry supports real-time inference with a managed, scalable REST endpoint?

  • A. Azure Batch Endpoints
  • B. Managed Online Endpoints✓ Correct
  • C. Azure Container Apps
  • D. Azure Functions Consumption Plan
Explanation

Managed Online Endpoints in Azure Machine Learning / Azure AI Foundry provide a fully managed, autoscalable REST endpoint for real-time inference. Models from the Foundry model catalog can be deployed directly to a managed online endpoint. Azure Batch Endpoints are designed for high-throughput offline batch scoring, not real-time inference. Azure Container Apps can host containerized workloads but is not a native deployment target in the Foundry model catalog workflow and requires manual containerization. Azure Functions on a Consumption Plan is serverless compute not integrated with the Foundry model catalog deployment flow, and it has cold-start and execution-duration constraints that are unsuitable for large model inference.

2. Woodgrove Bank deploys a GPT-4o model on Azure OpenAI Service for an internal knowledge assistant. Six months later, Microsoft announces that the specific model version (e.g., gpt-4o-2024-05-13) will reach its retirement date. A junior engineer suggests simply doing nothing because Azure will automatically upgrade the deployment. What is the CORRECT statement about Azure OpenAI model version retirement?

  • A. Azure OpenAI automatically upgrades the deployment to the next available model version on the retirement date with no action required
  • B. On the retirement date, the deployment is deleted and all API calls return HTTP 404 until manually redeployed
  • C. Customers must manually update their deployment to a newer model version before the retirement date, or the deployment will become unavailable✓ Correct
  • D. The model version remains available indefinitely as long as the customer's subscription is active
Explanation

Azure OpenAI model version retirement requires customers to take action — they must update their deployment to a supported model version before the retirement date. After retirement, the old version is no longer available for new calls. Automatic upgrades are available only when 'auto-update to default' is explicitly enabled on the deployment; it is not the default behavior for all deployments without configuration. The deployment is not deleted on retirement; rather, calls to the retired version begin to fail, but the resource itself is not removed in the same way as described. Model versions do not remain available indefinitely — Microsoft controls the lifecycle of hosted model versions.

3. A team is monitoring a production Azure OpenAI deployment using Azure Monitor. They have configured diagnostic settings to send logs to a Log Analytics workspace. A developer wants to write a KQL (Kusto Query Language) query to find all requests in the last 24 hours where the total token usage (prompt tokens + completion tokens) exceeded 3,000 tokens. Which Azure Monitor log table and field should the query target?

  • A. The 'AzureDiagnostics' table, filtering on the 'total_tokens_d' field emitted by the Azure OpenAI resource diagnostic logs✓ Correct
  • B. The 'AppRequests' table in Application Insights, filtering on the 'customDimensions.token_count' field
  • C. The 'AzureMetrics' table, filtering on the 'TokenTransaction' metric with a threshold greater than 3000
  • D. The 'ContainerLog' table, parsing JSON log lines emitted by the Azure OpenAI container runtime
Explanation

When Azure OpenAI diagnostic logs are sent to a Log Analytics workspace, they appear in the 'AzureDiagnostics' table. The per-request log entries include fields such as 'prompt_tokens_d', 'completion_tokens_d', and 'total_tokens_d' that can be queried with KQL to filter requests by token usage. 'AppRequests' is an Application Insights table for application-level telemetry and does not natively contain Azure OpenAI token usage fields unless custom instrumentation is added. 'AzureMetrics' contains aggregate metric time-series data (e.g., total tokens per minute) and does not provide per-request granularity needed to filter individual requests over 3,000 tokens. 'ContainerLog' contains stdout/stderr from containerized workloads and is unrelated to Azure OpenAI Service API diagnostic logging.

4. A solutions architect at Tailwind Traders is configuring Azure OpenAI on your data using an Azure AI Search index as the data source. Users are reporting that the chatbot sometimes answers questions with information that is NOT present in the company documents. Which TWO configuration changes are MOST likely to reduce this behavior? (Select TWO)

  • A. Set `in_scope: true` in the Azure OpenAI on your data API request body to restrict the model to only answer from the retrieved documents.✓ Correct
  • B. Increase the Azure OpenAI deployment's tokens-per-minute (TPM) quota to allow longer responses.
  • C. Enable semantic ranking on the Azure AI Search index so that the most relevant document chunks are retrieved.✓ Correct
  • D. Switch the Azure OpenAI model deployment from GPT-4o to GPT-4o-mini to reduce hallucination frequency.
  • E. Increase the `top` (number of retrieved documents) parameter to 50 to provide more context to the model.
Explanation

Options A and C are correct. Setting `in_scope: true` in the Azure OpenAI on your data payload instructs the model to only answer from retrieved documents and to respond with 'I don't know' if the answer isn't found — directly reducing out-of-scope answers. Enabling semantic ranking improves the relevance of retrieved chunks, ensuring the most pertinent information is passed to the model, which reduces the chance the model fabricates answers due to missing context. Option B is incorrect; TPM quota controls throughput, not answer grounding. Option D is incorrect; model size does not inherently reduce hallucinations — smaller models can hallucinate more. Option E is incorrect; passing 50 retrieved chunks often introduces noise and can actually increase hallucinations by diluting relevant content.

5. A DevOps engineer is managing Azure OpenAI deployments and needs to perform a zero-downtime model version upgrade for a GPT-4o deployment that currently serves 100% of production traffic. The team wants to test the new model version with 10% of live traffic before fully cutting over, and must be able to roll back instantly if issues arise. Which Azure capability and approach should they use?

  • A. Create a second Azure OpenAI resource in a different region with the new model version, then use Azure Front Door weighted routing to split traffic 90/10 between the two resources.
  • B. Deploy the new model version as a second deployment within the same Azure OpenAI resource, then use a managed online endpoint's traffic split feature to route 10% of requests to the new deployment and 90% to the existing one, reverting the split for rollback.
  • C. Use the Azure OpenAI Service's built-in A/B testing feature to configure a 90/10 traffic split between two model versions within a single deployment object.
  • D. Update the existing deployment's model version field in the Azure portal; Azure OpenAI will automatically perform a canary rollout at 10% traffic before completing the upgrade.
  • E. Deploy the new model version to a separate managed online endpoint and implement traffic split using the endpoint's deployment weight settings, using blue-green deployment for rollback.✓ Correct
Explanation

Option E is correct. For zero-downtime canary/blue-green model upgrades with fine-grained traffic control and instant rollback, the recommended approach is to deploy the new model version to a managed online endpoint (in Azure AI Foundry or Azure Machine Learning) and use the endpoint's deployment weight settings to split traffic (e.g., 10% to the new 'green' deployment, 90% to the existing 'blue' deployment). Rollback is instant by resetting weights to 100%/0%. Option A is incorrect; using Azure Front Door across two separate OpenAI resources is architecturally complex, harder to manage, and not the canonical approach—it also introduces cross-region latency inconsistencies. Option B is incorrect in its premise: Azure OpenAI Service itself does not provide native traffic splitting between two deployments within the same resource; that capability lives at the managed online endpoint layer. The option conflates OpenAI resource deployments with managed online endpoint traffic split. Option C is incorrect; there is no built-in A/B testing feature within a single Azure OpenAI deployment object—traffic split is not configurable at the deployment level. Option D is incorrect; Azure OpenAI does not automatically perform canary rollouts when you update a deployment's model version—it performs an in-place update.

160 more questions in this domain

Practice the full bank with instant grading, flashcards, and a timed mock exam.

Start practicing free