AI-103 practice questions

Question 11· Select all that apply

Implement generative AI and agentic solutions

Runs of a support agent sometimes finish with nothing added to the conversation. The team is working through the causes.

Which two would leave a run over with no reply written? Each correct selection is worth one point.

Select all that apply.

The run expired while waiting for a tool output the application never submitted.
The run failed, because the deployment behind the agent was returning errors.
The run was cancelled after the model had already finished writing its reply.
The agent holds more tools than the model can choose between in one turn.
The conversation was created before the agent was, so their identifiers differ.

Question 12· Yes or no

Plan and manage an Azure AI solution

A retrieval app runs in one region: a Foundry project with a chat deployment, a search index holding 400,000 passages, and the app itself behind a public address that customers have bookmarked. The business has asked that a regional outage cost no more than an hour of service.

The team has drawn a second region and has not built any of it yet. The exhibit shows the drawing as it stands, with the pieces that exist today on one side and the planned pieces on the other.

Two regions side by side: the region in production with its app, chat deployment and search index, and the planned region holding only a chat deployment, with client traffic arriving at the first

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

A deployment of the same model in the second region draws on that region's own allocation.
The index and its passages have to exist in the second region for the app there to answer.
Building the second region's deployment can wait until an outage, because capacity is granted within minutes of being asked for.
Customer traffic reaches the second region only if something in front of the app sends it there.

Question 13· Multiple choice

Implement generative AI and agentic solutions

A workflow scores loan applications. It runs a credit check, an address check and a fraud score, then calls an agent that writes the decision. None of the three checks reads the others' output.

Applicants complain the page takes too long. Every check has to keep running. What should you change?

Stream the decision so the applicant starts seeing text before the run ends.
Move the model to a provisioned deployment for a faster decision.
Put the three checks on parallel branches that join before the decision.
Cache the fraud score from the applicant's last visit.

Question 14· Yes or no

Plan and manage an Azure AI solution

An agent answers from supplier documents that suppliers upload themselves. The exhibit shows one request as the platform handles it.

One request through an agent, showing the prompt assembled from the system message, the user message and the uploaded supplier documents, then the input shields, the model, and the output filters

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

User prompt attack detection reads what the person typed.
An instruction hidden in an uploaded document is scored by the violence classifier.
Indirect attack detection needs the retrieved content marked as document content in the prompt.
Both shields report a severity between safe and high, like the harm categories.

Question 15· Multiple choice

Implement generative AI and agentic solutions

An assistant streams its replies. The default filtering mode holds the output back in short runs before releasing it, and the team can see the pause that puts in front of the first words. They move the deployment to the asynchronous filter, and the stream becomes smooth.

A reviewer asks what the change costs. What should you tell them?

A passage that would have been held back can reach the reader before its signal does.
The prompt stops being filtered, so only what the model writes is checked from here on.
Categories set to a high severity threshold stop being evaluated on streamed replies.
No text reaches the reader until the whole reply is checked, so the signal comes first.

Question 16· Complete the sentence

Plan and manage an Azure AI solution

An agent sends the same 12,000 token preamble of instructions and schema on every request, followed by a few hundred tokens of conversation. The team is sizing a provisioned deployment and has measured that the preamble is served from the prompt cache on nine calls in ten.

Complete the sizing note.

Input tokens served from the cacheprovisioned capacity, so the sizing calculation should.

Question 17· Complete the sentence

Implement generative AI and agentic solutions

A triage service reads three fields out of the model's reply with a regular expression. Whenever the model rephrases its answer the expression matches nothing and the request fails.

Complete the finding.

Ask for the reply. That guarantees.

Question 18· Yes or no

Implement computer vision solutions

A team is placing Content Safety image analysis in front of an upload form.

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

The image result carries a severity for each of the hate, sexual, violence and self-harm categories.
A blocklist attached to the request also matches text printed inside the image.
The severity at which an image is rejected is the caller's decision, per category.
The image has to be sent as a URL the service can fetch.

Question 19· Complete the sentence

Implement generative AI and agentic solutions

A prompt labels support messages by topic. The application stores the label. It is coming back as Billing, billing issue and BILLING for the same kind of message, and the store now holds three topics where there should be one.

Complete the prompt design.

Putin the system message, and.

Question 20· Multiple choice

Implement text analysis solutions

Reviews of a contoso coffee machine read like "the milk frother is superb but the app is hopeless". The product team does not want to know whether each review is positive. It wants a report naming the parts of the machine people praise and the parts they complain about.

What should the pipeline do?

Read the sentence-level sentiment, which scores each sentence on its own.
Run key phrase extraction, then run sentiment over each phrase it returns.
Run named entity recognition and count how often each entity appears.
Turn on opinion mining, which ties each assessment to what it is about.