McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 18

Question 170

A team promotes the same AI app from test to production. The model deployment name differs by environment, but the business logic should remain unchanged. What design minimizes deployment mistakes?

  1. Read deployment names and endpoints from environment configuration.
  2. Hard-code the test model name in the Python source for all environments.
  3. Change the model temperature whenever the environment changes to select another deployment.
  4. Use the Azure subscription display name as the model identifier without resolving the real deployment.

Explanation

Correct Answer

A:

Separating deployment identifiers from application code allows identical tested logic to target the approved environment.

Incorrect Answers

B:

A fixed test deployment name can route production requests to the wrong model or fail outright.

C:

Sampling controls do not choose an endpoint or deployment by name.

D:

A subscription label does not name an inference deployment.

Question 171

A chatbot must apply trusted business rules, read a user question, and consult a previous assistant response. Which input representation best preserves those distinct conversation roles?

  1. Concatenate all text into one anonymous string and rely on paragraph order as authorization.
  2. Set every message role to system, including untrusted user text.
  3. Put only the prior assistant response in the prompt and discard the current user request.
  4. Send system, user and assistant messages in their respective roles.

Explanation

Correct Answer

D:

Separating roles allows the model API to interpret instruction hierarchy and follow the conversation state correctly.

Incorrect Answers

A:

An undifferentiated string obscures the origin and role of messages, weakening intended instruction separation.

B:

Promoting user content to a system-level role changes the trust boundary and can create injection risk.

C:

The model would lack the current request it is supposed to answer.

Question 172

A chat assistant accumulates many turns of history. Soon requests fail because conversation context exceeds the deployed model limit, although the latest user prompt is short. Which response is appropriate?

  1. Increase the application concurrency setting; the model context limit will disappear.
  2. Maintain a bounded conversation window under context length limits.
  3. Return the entire chat twice to improve model recall.
  4. Disable all identity authorization checks to save context tokens.

Explanation

Correct Answer

B:

A managed conversation budget keeps prompt tokens within model limits while retaining relevant requirements.

Incorrect Answers

A:

Request parallelism does not enlarge the context window of an individual model call.

C:

Duplicating turns increases token pressure and can make the failure more likely.

D:

Removing authorization is unsafe and does not address the core excessive-history design.

Question 173

A tax-processing agent must return `country`, `tax_code`, and `confidence` to a strict API. Free-form prose repeatedly breaks parsing. Which output contract should the model use?

  1. Rely on the model to mention field names somewhere inside explanatory paragraphs.
  2. Generate a structured JSON response with an enforced output schema.
  3. Raise temperature to obtain more varied JSON structures.
  4. Parse the first three words of every response into the required columns.

Explanation

Correct Answer

B:

A strict schema reduces malformed integration payloads while explicit downstream validation remains necessary.

Incorrect Answers

A:

Prose mentions do not satisfy a machine-readable field contract.

C:

Variability makes a fixed-format contract harder to satisfy.

D:

Word position does not establish field semantics or types.

Question 174

A fulfillment service stores model-generated order details, but the downstream API rejects missing quantities and invalid SKUs. What should happen before persistence?

  1. Commit any response containing valid opening and closing braces without checking values.
  2. Validate typed JSON order fields before storage.
  3. Ask the model to promise the payload is valid and bypass the schema checker.
  4. Store the model response as a database comment and mark the order committed.

Explanation

Correct Answer

B:

Application-side validation prevents a syntactically plausible but incomplete order from entering the system.

Incorrect Answers

A:

JSON syntax alone does not prove required fields and business constraints are satisfied.

C:

A language assurance cannot replace deterministic validation.

D:

A comment is not an accepted and validated order transaction.

Question 175

A company chatbot must obey organizational instructions while handling untrusted customer requests. A developer proposes placing customers' text in the same system role as policy. Which design is correct?

  1. Store all messages as assistant-role output regardless of their origin.
  2. Ask users to provide the assistant's system instructions themselves each turn.
  3. Concatenate customer and policy instructions into an undifferentiated trusted system message.
  4. Keep trusted instructions in appropriate higher-priority roles and user content in its lower-trust role.

Explanation

Correct Answer

D:

Role separation prevents customer text from becoming implicit authority over application rules.

Incorrect Answers

A:

Incorrect role attribution loses the distinction between request, policy and prior response.

B:

User-provided instructions cannot define trusted enterprise policy.

C:

This promotes untrusted content to a privileged instruction position.

Question 176

Finance compares the cost of an inference call containing a large document but a very short answer. Which usage categories must be separated?

  1. Treat search index storage capacity as the direct per-call completion token count.
  2. Bill only the generated final answer characters and ignore document input.
  3. Use the number of chat participants as the total token count.
  4. Account for prompt and completion tokens at their respective rates.

Explanation

Correct Answer

D:

The two token directions contribute differently to usage and may have distinct prices.

Incorrect Answers

A:

Retrieval storage and model generation token usage are different measures.

B:

The substantial input document still consumes billable prompt tokens.

C:

Participant count does not quantify either input or output tokens.

Question 177

A document extractor returns different field wording for identical inputs, causing a compliance diff tool to report false changes. Which model setting should be evaluated first?

  1. Remove the output schema so the model can respond in any format.
  2. Increase request concurrency and assume the output becomes deterministic.
  3. Lower generation temperature and benchmark output consistency.
  4. Increase temperature to maximize creative diversity of extracted fields.

Explanation

Correct Answer

C:

Reduced temperature can improve repeatability, but correctness must still be measured.

Incorrect Answers

A:

Unconstrained formatting can increase rather than reduce false differences.

B:

Concurrency controls throughput rather than sampling consistency.

D:

More sampling variation conflicts with the request for stable extraction.

Question 178

A marketing team wants several creative but policy-compliant campaign themes rather than one near-identical wording. What generation adjustment may help?

  1. Turn off brand review because temperature determines compliance.
  2. Replace the generative model with an embedding-only search vector response.
  3. Force all runs to a minimal sampling setting and identical seed while demanding maximum variety.
  4. Choose a nonzero creativity setting for divergent marketing drafts.

Explanation

Correct Answer

D:

Moderate sampling variability can diversify concepts while safeguards remain necessary.

Incorrect Answers

A:

Sampling settings are not policy validation controls.

B:

Embeddings do not create the requested campaign prose.

C:

Low variability and fixed conditions can limit divergent output.

Question 179

An internal support bot can generate tens of pages after a trivial question, exhausting budget and UI capacity. Which request limit bounds the response?

  1. Bound maximum output tokens to prevent runaway long completions.
  2. Change only the sampling temperature without any length constraint.
  3. Increase the vector index dimension to shrink language answers.
  4. Use a higher search top-k as a cap on model completion tokens.

Explanation

Correct Answer

A:

The completion cap limits generation size and associated runaway output costs.

Incorrect Answers

B:

Sampling randomness is not a reliable hard output budget.

C:

Embedding dimensions do not set the generative response length.

D:

Retrieval candidate count does not directly bound the output size.

Question 180

An inference endpoint rejects a harmful user request with a content-filter outcome. A developer plans to return the rejected raw prompt in a detailed error page. What should the app do?

  1. Retry the same blocked request indefinitely until one attempt gets through.
  2. Handle content filter responses without returning unsanitized raw text.
  3. Disable moderation for users whose initial request was rejected.
  4. Treat the filter error as a successful generated answer and display its entire payload.

Explanation

Correct Answer

B:

Safety interventions should be surfaced without reintroducing prohibited material through the UI.

Incorrect Answers

A:

Unlimited retries undermine policy enforcement and waste resources.

C:

A rejection is not authorization to remove safety controls.

D:

This can bypass the intended safety treatment of the rejected content.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests