AI-103 practice questions

Question 1· Yes or no

Implement generative AI and agentic solutions

The exhibit shows four runs of the same agent, with the tokens each one spent and how long it took.

Token usage for four agent runs, with input tokens, output tokens, tool calls and duration for each

For each statement, select Yes if it is true. Otherwise, select No.

The run that produced the most output tokens is the slowest of the four.
Input tokens climb as a conversation goes on, because earlier turns are sent again.
Token counts have to be logged by the application, because a trace does not carry them.
Two runs answering the same question can differ in input tokens if one retrieved more passages.

Question 2· Multiple choice

Plan and manage an Azure AI solution

App1 runs on a virtual machine in a virtual network and calls a Foundry project. You create a private endpoint for the project, turn off public network access on it, and leave the virtual machine using the default Azure resolver.

App1 can no longer connect, and a name lookup for the project from the virtual machine returns a public address.

The virtual network holding the virtual machine and the private endpoint, the Foundry project, and a private DNS zone with its record and its link count

What is missing?

A route in the virtual network for the project address range.
An outbound rule letting the virtual machine reach the endpoint.
A link between the private DNS zone and the virtual network.
A second record in the private DNS zone, for the regional name.

Question 3· Put in order

Implement generative AI and agentic solutions

A team wants to know whether a rewritten prompt is better than the one in production, rather than whether it feels better.

Put the steps in the order they run.

1Score the recorded rows with the evaluators the team agreed on.
2Assemble a dataset of questions the assistant has to handle.
3Ship the rewrite or send it back for another revision.
4Run the assistant over the dataset and record each answer with its retrieved passages.
5Compare the scores against the run the current prompt produced.

Question 4· Complete the sentence

Plan and manage an Azure AI solution

A developer assistant writes code snippets for an engineering team. Legal has asked that when a snippet matches source from a public repository, the engineer is told where it came from. They have also asked that the assistant keep answering, because an assistant that goes silent is one nobody uses.

Complete the configuration.

Set protected material for code to, and have the application.

Question 5· Put in order

Implement generative AI and agentic solutions

The exhibit shows the objects a Foundry agent run involves and how they relate to each other.

An agent definition, a thread holding messages, a run naming an agent, and the calling application

An agent with one function tool already exists. Put the steps the application takes in the order they happen for one question.

1Add the user's question to the thread as a message.
2Create a thread for the conversation.
3Read the agent's reply from the thread.
4Start a run on the thread, naming the agent.
5Submit the output of the function the run asked for.

Question 6· Select all that apply

Plan and manage an Azure AI solution

A pipeline creates a model deployment in a Foundry project in North Europe. The step fails. The same template creates the same deployment in West Europe without complaint, and the project itself is healthy.

Which two things would explain it? Each correct selection is worth one point.

Select all that apply.

The subscription has no allocation left for that model in North Europe.
The model is not offered in North Europe.
The project in North Europe has no connection to a search service.
No diagnostic setting has been added to the North Europe account.
The deployment's rate limit was left at its default value.

Question 7· Multiple choice

Implement generative AI and agentic solutions

Users report that an agent takes about nine seconds to answer. The project is connected to an Application Insights resource, and the exhibit shows the spans recorded for one slow run.

A trace for one agent run, showing the run span and the child spans for each model call and each tool call with their durations

The team plans to move the model to a provisioned deployment to bring the nine seconds down. What does the trace say about that plan?

Most of the run sat in one tool call, so a faster model buys back very little.
The model calls take most of the run, so a provisioned deployment is the right move.
The trace cannot separate model time from tool time, so more instrumentation is needed.
The tools produced no spans, so client-side tracing has to be added before deciding.

Question 8· Multiple choice

Plan and manage an Azure AI solution

A customer complained that an order agent told them their parcel was in transit when it had been delivered two days earlier. The exhibit shows that run.

One agent run, showing the user question, a single call to the order lookup tool, the result it returned saying delivered, and the reply the agent sent saying in transit

Which evaluator reports this run as a failure?

Tool call success
Tool selection
Intent resolution
Tool output utilization

Question 9· Multiple choice

Implement generative AI and agentic solutions

An agent has two tools. lookup_invoice returns an invoice by number and lookup_account returns a customer account. Both descriptions read "Looks up a record."

The agent calls the wrong one about half the time. What should you change first?

The tool descriptions, so each says what it returns and when to use it.
The agent instructions, so they say what each tool returns and when.
The model deployment, so a stronger model tells the two tools apart.
The tool order, so the more common description is offered to the model first.

Question 10· Complete the sentence

Plan and manage an Azure AI solution

A search service backs a retrieval app. At peak, queries queue and time out while the index sits at a third of the storage the service allows. Six months later the corpus has tripled and the service reports the index close to its storage ceiling.

The exhibit shows the scale settings the service is running today.

Complete the plan.

For the queries timing out at peak, add. For the tripled corpus six months later, add.