Practice Exams:

Microsoft AI-103 Dumps

EXAM: Developing AI Apps and Agents on Azure
Verified By Experts
Premium Bundle
$49.99

AI-103 Premium ETE File


  • Premium File 187 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-103 Practice Test 17

Question 160

During a production incident, a cloud developer must accept a change for token cost anomaly. The issue to be addressed is: Overnight agent runs begin consuming far more output tokens than usual. Which concrete verification action demonstrates that the relevant behavior works?

  1. Compare traced prompt and model versions for an observed regression
  2. Inspect per-run prompt and completion token telemetry and isolate new expensive flows
  3. Judge cost only from the number of API requests while ignoring generated length
  4. Compare total tokens and spend before and after a traced optimization

Explanation

Correct Answer

D:

To verify this change, the education technology company should compare total tokens and spend before and after a traced optimization. Operators still need the implementation to inspect per-run prompt and completion token telemetry and isolate new expensive flows and the preventive requirement to set per-task token ceilings and alert on cost spikes.

Incorrect Answers

A:

For experiment reproducibility, it can be sensible to compare traced prompt and model versions for an observed regression if a team cannot explain which prompt version produced a degraded answer. Without the actual change to compare total tokens and spend before and after a traced optimization, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

B:

The action to inspect per-run prompt and completion token telemetry and isolate new expensive flows can contribute to token cost anomaly as a implementation step; here, overnight agent runs begin consuming far more output tokens than usual. In the education technology company acceptance review, trying to inspect per-run prompt and completion token telemetry and isolate new expensive flows does not test the exact token cost anomaly condition where overnight agent runs begin consuming far more output tokens than usual; the relevant test is to compare total tokens and spend before and after a traced optimization.

C:

In the education technology company acceptance review, trying to judge cost only from the number of API requests while ignoring generated length does not test the exact token cost anomaly condition where overnight agent runs begin consuming far more output tokens than usual; the relevant test is to compare total tokens and spend before and after a traced optimization. Acceptance should instead demonstrate that the service can compare total tokens and spend before and after a traced optimization, using the described failure as the test criterion.

Question 161

At a public library network, the site reliability engineer is preparing a architecture migration. The concrete requirement is: An otherwise fast chatbot responds slowly after retrieving external data. Which design action BEST addresses the stated need?

  1. Restrict retrieval connector data-plane permissions to authorized corpora
  2. Use distributed traces to separate model time from individual tool execution latency
  3. Blame model decoding speed without observing external tool call durations
  4. Inspect a slow trace and identify the tool step consuming the delay

Explanation

Correct Answer

B:

The proposed implementation is to use distributed traces to separate model time from individual tool execution latency; this matches the operational requirement that an otherwise fast chatbot responds slowly after retrieving external data. The team can inspect a slow trace and identify the tool step consuming the delay. The team should also instrument each tool invocation with latency and timeout metrics.

Incorrect Answers

A:

Choosing to restrict retrieval connector data-plane permissions to authorized corpora cannot directly replace the specified action to use distributed traces to separate model time from individual tool execution latency when an otherwise fast chatbot responds slowly after retrieving external data. The relevant corrective or preventive action here remains to use distributed traces to separate model time from individual tool execution latency, with its own validation criteria.

C:

Choosing to blame model decoding speed without observing external tool call durations cannot directly replace the specified action to use distributed traces to separate model time from individual tool execution latency when an otherwise fast chatbot responds slowly after retrieving external data. The actual implementation must use distributed traces to separate model time from individual tool execution latency, because the stated service condition is that an otherwise fast chatbot responds slowly after retrieving external data.

D:

The action to inspect a slow trace and identify the tool step consuming the delay can contribute to agent tool latency as a targeted acceptance check step; here, an otherwise fast chatbot responds slowly after retrieving external data. Choosing to inspect a slow trace and identify the tool step consuming the delay cannot directly replace the specified action to use distributed traces to separate model time from individual tool execution latency when an otherwise fast chatbot responds slowly after retrieving external data.

Question 162

After a operational readiness review at a software consultancy, the team observes this failure: An otherwise fast chatbot responds slowly after retrieving external data. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Blame model decoding speed without observing external tool call durations
  2. Inspect a slow trace and identify the tool step consuming the delay
  3. Assume application authentication automatically restricts administrator permissions
  4. Use distributed traces to separate model time from individual tool execution latency

Explanation

Correct Answer

A:

This design mistake is to blame model decoding speed without observing external tool call durations; it conflicts with the actual agent tool latency behavior needed when an otherwise fast chatbot responds slowly after retrieving external data. The software consultancy should instead use distributed traces to separate model time from individual tool execution latency, followed by a check to inspect a slow trace and identify the tool step consuming the delay.

Incorrect Answers

B:

The alternative to inspect a slow trace and identify the tool step consuming the delay is a targeted acceptance check step rather than the faulty assumption being diagnosed. This software consultancy failure-analysis alternative of inspect a slow trace and identify the tool step consuming the delay does not diagnose the particular agent tool latency defect where an otherwise fast chatbot responds slowly after retrieving external data; the keyed mistake is blame model decoding speed without observing external tool call durations.

C:

This software consultancy failure-analysis alternative of assume application authentication automatically restricts administrator permissions does not diagnose the particular agent tool latency defect where an otherwise fast chatbot responds slowly after retrieving external data; the keyed mistake is blame model decoding speed without observing external tool call durations. The alternative may be useful in its own context, but it does not directly supply blame model decoding speed without observing external tool call durations for the stated decision.

D:

The alternative to use distributed traces to separate model time from individual tool execution latency is a implementation step rather than the faulty assumption being diagnosed. This software consultancy failure-analysis alternative of use distributed traces to separate model time from individual tool execution latency does not diagnose the particular agent tool latency defect where an otherwise fast chatbot responds slowly after retrieving external data; the keyed mistake is blame model decoding speed without observing external tool call durations.

Question 163

The regional hospital is responding to a regional deployment. A agent designer discovers this issue: An otherwise fast chatbot responds slowly after retrieving external data. What corrective step should the team take to restore the intended functionality most directly?

  1. Blame model decoding speed without observing external tool call durations
  2. Inspect a slow trace and identify the tool step consuming the delay
  3. Improve retriever ranking and replay known grounded answers
  4. Optimize the slow tool or apply bounded timeouts and safe fallbacks

Explanation

Correct Answer

D:

The direct repair is to optimize the slow tool or apply bounded timeouts and safe fallbacks. Once the condition is corrected, the regional hospital should inspect a slow trace and identify the tool step consuming the delay. To reduce recurrence, it should also instrument each tool invocation with latency and timeout metrics; those steps provide different recovery and prevention functions.

Incorrect Answers

A:

During regional hospital remediation, the proposed step to blame model decoding speed without observing external tool call durations is not the direct restoration action for agent tool latency when an otherwise fast chatbot responds slowly after retrieving external data; the keyed repair is optimize the slow tool or apply bounded timeouts and safe fallbacks. After remediation the team must inspect a slow trace and identify the tool step consuming the delay; the action in this option does not itself establish that target recovery outcome.

B:

The action to inspect a slow trace and identify the tool step consuming the delay can contribute to agent tool latency as a targeted acceptance check step; here, an otherwise fast chatbot responds slowly after retrieving external data. During regional hospital remediation, the proposed step to inspect a slow trace and identify the tool step consuming the delay is not the direct restoration action for agent tool latency when an otherwise fast chatbot responds slowly after retrieving external data; the keyed repair is optimize the slow tool or apply bounded timeouts and safe fallbacks.

C:

This approach relates to grounding quality monitoring: a RAG bot returns fluent answers that cite irrelevant source passages. If the team chose it instead, the targeted check to inspect a slow trace and identify the tool step consuming the delay could still fail because the underlying required action is to optimize the slow tool or apply bounded timeouts and safe fallbacks.

Question 164

A responsible-AI reviewer at a insurance platform wants the following issue to be prevented in later deployments: An otherwise fast chatbot responds slowly after retrieving external data. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Define relevance and evidence-support thresholds with human audit samples
  2. Instrument each tool invocation with latency and timeout metrics
  3. Blame model decoding speed without observing external tool call durations
  4. Optimize the slow tool or apply bounded timeouts and safe fallbacks

Explanation

Correct Answer

B:

Standardizing the boundary to instrument each tool invocation with latency and timeout metrics is preventive because the same agent tool latency failure could otherwise return: an otherwise fast chatbot responds slowly after retrieving external data. For this insurance platform, the service should also use distributed traces to separate model time from individual tool execution latency, and tests should inspect a slow trace and identify the tool step consuming the delay so that the ongoing requirement is measurable.

Incorrect Answers

A:

The action to define relevance and evidence-support thresholds with human audit samples addresses grounding quality monitoring rather than the decisive agent tool latency condition. The insurance platform should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to instrument each tool invocation with latency and timeout metrics.

C:

For prevention at the insurance platform, the measure to blame model decoding speed without observing external tool call durations is weaker or aimed at another issue than the agent tool latency guardrail to instrument each tool invocation with latency and timeout metrics needed when an otherwise fast chatbot responds slowly after retrieving external data. The lasting guardrail for this specific threat is to instrument each tool invocation with latency and timeout metrics, not simply to complete an adjacent service task.

D:

The action to optimize the slow tool or apply bounded timeouts and safe fallbacks can contribute to agent tool latency as a corrective remediation step; here, an otherwise fast chatbot responds slowly after retrieving external data. For prevention at the insurance platform, the measure to optimize the slow tool or apply bounded timeouts and safe fallbacks is weaker or aimed at another issue than the agent tool latency guardrail to instrument each tool invocation with latency and timeout metrics needed when an otherwise fast chatbot responds slowly after retrieving external data.

Question 165

During a operational readiness review, a solution architect must accept a change for agent tool latency. The issue to be addressed is: An otherwise fast chatbot responds slowly after retrieving external data. Which concrete verification action demonstrates that the relevant behavior works?

  1. Use distributed traces to separate model time from individual tool execution latency
  2. Invoke the app under its actual service principal rather than a local user
  3. Instrument each tool invocation with latency and timeout metrics
  4. Inspect a slow trace and identify the tool step consuming the delay
  5. Blame model decoding speed without observing external tool call durations

Explanation

Correct Answer

D:

The targeted acceptance test is to inspect a slow trace and identify the tool step consuming the delay. The design should first use distributed traces to separate model time from individual tool execution latency. A robust media platform deployment also needs to instrument each tool invocation with latency and timeout metrics; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

The action to use distributed traces to separate model time from individual tool execution latency can contribute to agent tool latency as a implementation step; here, an otherwise fast chatbot responds slowly after retrieving external data. In the media platform acceptance review, trying to use distributed traces to separate model time from individual tool execution latency does not test the exact agent tool latency condition where an otherwise fast chatbot responds slowly after retrieving external data; the relevant test is to inspect a slow trace and identify the tool step consuming the delay.

B:

Choosing to invoke the app under its actual service principal rather than a local user would direct effort toward access token authorization, whose characteristic issue is that an inference app succeeds locally using developer credentials but fails under deployment. The relevant corrective or preventive action here remains to inspect a slow trace and identify the tool step consuming the delay, with its own validation criteria.

C:

The action to instrument each tool invocation with latency and timeout metrics can contribute to agent tool latency as a ongoing preventive rule step; here, an otherwise fast chatbot responds slowly after retrieving external data. In the media platform acceptance review, trying to instrument each tool invocation with latency and timeout metrics does not test the exact agent tool latency condition where an otherwise fast chatbot responds slowly after retrieving external data; the relevant test is to inspect a slow trace and identify the tool step consuming the delay.

E:

In the media platform acceptance review, trying to blame model decoding speed without observing external tool call durations does not test the exact agent tool latency condition where an otherwise fast chatbot responds slowly after retrieving external data; the relevant test is to inspect a slow trace and identify the tool step consuming the delay. Acceptance should instead demonstrate that the service can inspect a slow trace and identify the tool step consuming the delay, using the described failure as the test criterion.

Question 166

During a cost review at a education technology company, an AI team must address grounding quality monitoring. The requirement is: A RAG bot returns fluent answers that cite irrelevant source passages. Which THREE actions together implement the needed solution, its preventive guardrail, and a targeted acceptance test? (Select THREE answers.)

  1. Improve retriever ranking and replay known grounded answers
  2. Evaluate retrieval relevance and groundedness separately from answer fluency
  3. Define relevance and evidence-support thresholds with human audit samples
  4. Define service-specific health SLOs and dependency diagrams
  5. Assume citations prove factual support without checking cited text
  6. Compare cited chunks to a labeled answer-evidence evaluation set

Explanation

Correct Answers

B:

Choosing to evaluate retrieval relevance and groundedness separately from answer fluency directly targets the described grounding quality monitoring gap: a RAG bot returns fluent answers that cite irrelevant source passages. As a lasting boundary, they also need to define relevance and evidence-support thresholds with human audit samples, because an implementation by itself does not prove safe operation.

C:

Standardizing the boundary to define relevance and evidence-support thresholds with human audit samples is preventive because the same grounding quality monitoring failure could otherwise return: a RAG bot returns fluent answers that cite irrelevant source passages. For this education technology company, the service should also evaluate retrieval relevance and groundedness separately from answer fluency, and tests should compare cited chunks to a labeled answer-evidence evaluation set so that the ongoing requirement is measurable.

F:

To verify this change, the education technology company should compare cited chunks to a labeled answer-evidence evaluation set. Operators still need the implementation to evaluate retrieval relevance and groundedness separately from answer fluency and the preventive requirement to define relevance and evidence-support thresholds with human audit samples.

Incorrect Answers

A:

The action to improve retriever ranking and replay known grounded answers can contribute to grounding quality monitoring as a corrective remediation step; here, a RAG bot returns fluent answers that cite irrelevant source passages. Choosing to improve retriever ranking and replay known grounded answers cannot directly replace the specified action to evaluate retrieval relevance and groundedness separately from answer fluency when a RAG bot returns fluent answers that cite irrelevant source passages.

D:

This approach relates to provider health monitoring: an app needs to distinguish model failures from search and speech outages. If the team chose it instead, the targeted check to compare cited chunks to a labeled answer-evidence evaluation set could still fail because the underlying required action is to evaluate retrieval relevance and groundedness separately from answer fluency.

E:

Choosing to assume citations prove factual support without checking cited text cannot directly replace the specified action to evaluate retrieval relevance and groundedness separately from answer fluency when a RAG bot returns fluent answers that cite irrelevant source passages. The actual implementation must evaluate retrieval relevance and groundedness separately from answer fluency, because the stated service condition is that a RAG bot returns fluent answers that cite irrelevant source passages.

Question 167

After a reliability exercise at a pharmaceutical research group, the team observes this failure: A RAG bot returns fluent answers that cite irrelevant source passages. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Enable full prompt-body logging in a broadly readable workspace by default
  2. Compare cited chunks to a labeled answer-evidence evaluation set
  3. Assume citations prove factual support without checking cited text
  4. Evaluate retrieval relevance and groundedness separately from answer fluency

Explanation

Correct Answer

C:

The faulty assumption is assume citations prove factual support without checking cited text. A functioning alternative would be to evaluate retrieval relevance and groundedness separately from answer fluency, and the relevant validation would be to compare cited chunks to a labeled answer-evidence evaluation set.

Incorrect Answers

A:

This pharmaceutical research group failure-analysis alternative of enable full prompt-body logging in a broadly readable workspace by default does not diagnose the particular grounding quality monitoring defect where a RAG bot returns fluent answers that cite irrelevant source passages; the keyed mistake is assume citations prove factual support without checking cited text. The alternative may be useful in its own context, but it does not directly supply assume citations prove factual support without checking cited text for the stated decision.

B:

The alternative to compare cited chunks to a labeled answer-evidence evaluation set is a targeted acceptance check step rather than the faulty assumption being diagnosed. This pharmaceutical research group failure-analysis alternative of compare cited chunks to a labeled answer-evidence evaluation set does not diagnose the particular grounding quality monitoring defect where a RAG bot returns fluent answers that cite irrelevant source passages; the keyed mistake is assume citations prove factual support without checking cited text.

D:

The alternative to evaluate retrieval relevance and groundedness separately from answer fluency is a implementation step rather than the faulty assumption being diagnosed. This pharmaceutical research group failure-analysis alternative of evaluate retrieval relevance and groundedness separately from answer fluency does not diagnose the particular grounding quality monitoring defect where a RAG bot returns fluent answers that cite irrelevant source passages; the keyed mistake is assume citations prove factual support without checking cited text.

Question 168

The travel support provider is responding to a source update. A quality assurance lead discovers this issue: A RAG bot returns fluent answers that cite irrelevant source passages. What corrective step should the team take to restore the intended functionality most directly?

  1. Improve retriever ranking and replay known grounded answers
  2. Compare cited chunks to a labeled answer-evidence evaluation set
  3. Assume citations prove factual support without checking cited text
  4. Remove overbroad role grants and retest required developer workflows

Explanation

Correct Answer

A:

The appropriate remediation is to improve retriever ranking and replay known grounded answers, because the failure concerns grounding quality monitoring: a RAG bot returns fluent answers that cite irrelevant source passages. A meaningful verification should compare cited chunks to a labeled answer-evidence evaluation set, and the durable operating rule should define relevance and evidence-support thresholds with human audit samples after the immediate fix.

Incorrect Answers

B:

The action to compare cited chunks to a labeled answer-evidence evaluation set can contribute to grounding quality monitoring as a targeted acceptance check step; here, a RAG bot returns fluent answers that cite irrelevant source passages. During travel support provider remediation, the proposed step to compare cited chunks to a labeled answer-evidence evaluation set is not the direct restoration action for grounding quality monitoring when a RAG bot returns fluent answers that cite irrelevant source passages; the keyed repair is improve retriever ranking and replay known grounded answers.

C:

During travel support provider remediation, the proposed step to assume citations prove factual support without checking cited text is not the direct restoration action for grounding quality monitoring when a RAG bot returns fluent answers that cite irrelevant source passages; the keyed repair is improve retriever ranking and replay known grounded answers. After remediation the team must compare cited chunks to a labeled answer-evidence evaluation set; the action in this option does not itself establish that target recovery outcome.

D:

During travel support provider remediation, the proposed step to remove overbroad role grants and retest required developer workflows is not the direct restoration action for grounding quality monitoring when a RAG bot returns fluent answers that cite irrelevant source passages; the keyed repair is improve retriever ranking and replay known grounded answers. If the team chose it instead, the targeted check to compare cited chunks to a labeled answer-evidence evaluation set could still fail because the underlying required action is to improve retriever ranking and replay known grounded answers.

Question 169

A AI platform engineer at a scientific publisher wants the following issue to be prevented in later deployments: A RAG bot returns fluent answers that cite irrelevant source passages. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Define relevance and evidence-support thresholds with human audit samples
  2. Assume citations prove factual support without checking cited text
  3. Improve retriever ranking and replay known grounded answers
  4. Alert on ingestion lag and reject stale high-risk policy evidence

Explanation

Correct Answer

A:

The sustained prevention measure is to define relevance and evidence-support thresholds with human audit samples. The implementation can still use the separate action to evaluate retrieval relevance and groundedness separately from answer fluency, but a guardrail constrains future changes. The scientific publisher can measure whether the rule works by arranging to compare cited chunks to a labeled answer-evidence evaluation set.

Incorrect Answers

B:

For prevention at the scientific publisher, the measure to assume citations prove factual support without checking cited text is weaker or aimed at another issue than the grounding quality monitoring guardrail to define relevance and evidence-support thresholds with human audit samples needed when a RAG bot returns fluent answers that cite irrelevant source passages. The lasting guardrail for this specific threat is to define relevance and evidence-support thresholds with human audit samples, not simply to complete an adjacent service task.

C:

The action to improve retriever ranking and replay known grounded answers can contribute to grounding quality monitoring as a corrective remediation step; here, a RAG bot returns fluent answers that cite irrelevant source passages. For prevention at the scientific publisher, the measure to improve retriever ranking and replay known grounded answers is weaker or aimed at another issue than the grounding quality monitoring guardrail to define relevance and evidence-support thresholds with human audit samples needed when a RAG bot returns fluent answers that cite irrelevant source passages.

D:

The action to alert on ingestion lag and reject stale high-risk policy evidence addresses search index health rather than the decisive grounding quality monitoring condition. The scientific publisher should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to define relevance and evidence-support thresholds with human audit samples.

Question 170

During a reliability exercise, a compliance specialist must accept a change for grounding quality monitoring. The issue to be addressed is: A RAG bot returns fluent answers that cite irrelevant source passages. Which concrete verification action demonstrates that the relevant behavior works?

  1. Compare cited chunks to a labeled answer-evidence evaluation set
  2. Verify developers can deploy only permitted resources and cannot change unrelated ones
  3. Assume citations prove factual support without checking cited text
  4. Evaluate retrieval relevance and groundedness separately from answer fluency
  5. Define relevance and evidence-support thresholds with human audit samples
  6. Improve retriever ranking and replay known grounded answers

Explanation

Correct Answer

A:

To verify this change, the nonprofit organization should compare cited chunks to a labeled answer-evidence evaluation set. Operators still need the implementation to evaluate retrieval relevance and groundedness separately from answer fluency and the preventive requirement to define relevance and evidence-support thresholds with human audit samples.

Incorrect Answers

B:

The option to verify developers can deploy only permitted resources and cannot change unrelated ones would normally serve a different azure role scope requirement, where a developer can unexpectedly edit all AI project resources. The alternative may be useful in its own context, but it does not directly supply compare cited chunks to a labeled answer-evidence evaluation set for the stated decision.

C:

In the nonprofit organization acceptance review, trying to assume citations prove factual support without checking cited text does not test the exact grounding quality monitoring condition where a RAG bot returns fluent answers that cite irrelevant source passages; the relevant test is to compare cited chunks to a labeled answer-evidence evaluation set. Acceptance should instead demonstrate that the service can compare cited chunks to a labeled answer-evidence evaluation set, using the described failure as the test criterion.

D:

The action to evaluate retrieval relevance and groundedness separately from answer fluency can contribute to grounding quality monitoring as a implementation step; here, a RAG bot returns fluent answers that cite irrelevant source passages. In the nonprofit organization acceptance review, trying to evaluate retrieval relevance and groundedness separately from answer fluency does not test the exact grounding quality monitoring condition where a RAG bot returns fluent answers that cite irrelevant source passages; the relevant test is to compare cited chunks to a labeled answer-evidence evaluation set.

E:

The action to define relevance and evidence-support thresholds with human audit samples can contribute to grounding quality monitoring as a ongoing preventive rule step; here, a RAG bot returns fluent answers that cite irrelevant source passages. In the nonprofit organization acceptance review, trying to define relevance and evidence-support thresholds with human audit samples does not test the exact grounding quality monitoring condition where a RAG bot returns fluent answers that cite irrelevant source passages; the relevant test is to compare cited chunks to a labeled answer-evidence evaluation set.

F:

The action to improve retriever ranking and replay known grounded answers can contribute to grounding quality monitoring as a corrective remediation step; here, a RAG bot returns fluent answers that cite irrelevant source passages. In the nonprofit organization acceptance review, trying to improve retriever ranking and replay known grounded answers does not test the exact grounding quality monitoring condition where a RAG bot returns fluent answers that cite irrelevant source passages; the relevant test is to compare cited chunks to a labeled answer-evidence evaluation set.

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