Microsoft AI-103 Dumps
Microsoft AI-103 Practice Test 18
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?
Explanation
Correct Answer
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
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.
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.
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.
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.
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.
Question 171
At a media platform, the product support lead is preparing a service audit. The concrete requirement is: An assistant becomes less accurate after a deployment or prompt change. Which design action BEST addresses the stated need?
Explanation
Correct Answer
The proposed implementation is to compare versioned evaluation scores against a prechange baseline and relevant user cohorts; this matches the operational requirement that an assistant becomes less accurate after a deployment or prompt change. The team can evaluate a fixed representative prompt set across both versions. The team should also require a stable regression test suite and deployment annotations.
Incorrect Answers
Choosing to restrict retrieval connector data-plane permissions to authorized corpora cannot directly replace the specified action to compare versioned evaluation scores against a prechange baseline and relevant user cohorts when an assistant becomes less accurate after a deployment or prompt change. The relevant corrective or preventive action here remains to compare versioned evaluation scores against a prechange baseline and relevant user cohorts, with its own validation criteria.
The action to evaluate a fixed representative prompt set across both versions can contribute to model output drift as a targeted acceptance check step; here, an assistant becomes less accurate after a deployment or prompt change. Choosing to evaluate a fixed representative prompt set across both versions cannot directly replace the specified action to compare versioned evaluation scores against a prechange baseline and relevant user cohorts when an assistant becomes less accurate after a deployment or prompt change.
Choosing to treat reduced complaint volume as definitive evidence of unchanged model quality cannot directly replace the specified action to compare versioned evaluation scores against a prechange baseline and relevant user cohorts when an assistant becomes less accurate after a deployment or prompt change. The actual implementation must compare versioned evaluation scores against a prechange baseline and relevant user cohorts, because the stated service condition is that an assistant becomes less accurate after a deployment or prompt change.
Question 172
After a security tabletop exercise at a customer support center, the team observes this failure: An assistant becomes less accurate after a deployment or prompt change. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
This design mistake is to treat reduced complaint volume as definitive evidence of unchanged model quality; it conflicts with the actual model output drift behavior needed when an assistant becomes less accurate after a deployment or prompt change. The customer support center should instead compare versioned evaluation scores against a prechange baseline and relevant user cohorts, followed by a check to evaluate a fixed representative prompt set across both versions.
Incorrect Answers
The option to depend solely on a single resource-state indicator to prove request success would normally serve a different deployment availability requirement, where a regional service has intermittent errors but its health endpoint reports green. The alternative may be useful in its own context, but it does not directly supply treat reduced complaint volume as definitive evidence of unchanged model quality for the stated decision.
The alternative to compare versioned evaluation scores against a prechange baseline and relevant user cohorts is a implementation step rather than the faulty assumption being diagnosed. This customer support center failure-analysis alternative of compare versioned evaluation scores against a prechange baseline and relevant user cohorts does not diagnose the particular model output drift defect where an assistant becomes less accurate after a deployment or prompt change; the keyed mistake is treat reduced complaint volume as definitive evidence of unchanged model quality.
The alternative to evaluate a fixed representative prompt set across both versions is a targeted acceptance check step rather than the faulty assumption being diagnosed. This customer support center failure-analysis alternative of evaluate a fixed representative prompt set across both versions does not diagnose the particular model output drift defect where an assistant becomes less accurate after a deployment or prompt change; the keyed mistake is treat reduced complaint volume as definitive evidence of unchanged model quality.
Question 173
The manufacturing group is responding to a support escalation. A service administrator discovers this issue: An assistant becomes less accurate after a deployment or prompt change. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The direct repair is to roll back the changed deployment when regression crosses an approved threshold. Once the condition is corrected, the manufacturing group should evaluate a fixed representative prompt set across both versions. To reduce recurrence, it should also require a stable regression test suite and deployment annotations; those steps provide different recovery and prevention functions.
Incorrect Answers
This approach relates to private data telemetry: model traces risk capturing customer health details in plaintext. If the team chose it instead, the targeted check to evaluate a fixed representative prompt set across both versions could still fail because the underlying required action is to roll back the changed deployment when regression crosses an approved threshold.
The action to evaluate a fixed representative prompt set across both versions can contribute to model output drift as a targeted acceptance check step; here, an assistant becomes less accurate after a deployment or prompt change. During manufacturing group remediation, the proposed step to evaluate a fixed representative prompt set across both versions is not the direct restoration action for model output drift when an assistant becomes less accurate after a deployment or prompt change; the keyed repair is roll back the changed deployment when regression crosses an approved threshold.
During manufacturing group remediation, the proposed step to treat reduced complaint volume as definitive evidence of unchanged model quality is not the direct restoration action for model output drift when an assistant becomes less accurate after a deployment or prompt change; the keyed repair is roll back the changed deployment when regression crosses an approved threshold. After remediation the team must evaluate a fixed representative prompt set across both versions; the action in this option does not itself establish that target recovery outcome.
Question 174
A cloud developer at a legal operations unit wants the following issue to be prevented in later deployments: An assistant becomes less accurate after a deployment or prompt change. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
Standardizing the boundary to require a stable regression test suite and deployment annotations is preventive because the same model output drift failure could otherwise return: an assistant becomes less accurate after a deployment or prompt change. For this legal operations unit, the service should also compare versioned evaluation scores against a prechange baseline and relevant user cohorts, and tests should evaluate a fixed representative prompt set across both versions so that the ongoing requirement is measurable.
Incorrect Answers
The action to roll back the changed deployment when regression crosses an approved threshold can contribute to model output drift as a corrective remediation step; here, an assistant becomes less accurate after a deployment or prompt change. For prevention at the legal operations unit, the measure to roll back the changed deployment when regression crosses an approved threshold is weaker or aimed at another issue than the model output drift guardrail to require a stable regression test suite and deployment annotations needed when an assistant becomes less accurate after a deployment or prompt change.
The action to perform periodic privileged access reviews for AI project resources addresses azure role scope rather than the decisive model output drift condition. The legal operations unit should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to require a stable regression test suite and deployment annotations.
For prevention at the legal operations unit, the measure to treat reduced complaint volume as definitive evidence of unchanged model quality is weaker or aimed at another issue than the model output drift guardrail to require a stable regression test suite and deployment annotations needed when an assistant becomes less accurate after a deployment or prompt change. The lasting guardrail for this specific threat is to require a stable regression test suite and deployment annotations, not simply to complete an adjacent service task.
Question 175
During a security tabletop exercise, a site reliability engineer must accept a change for model output drift. The issue to be addressed is: An assistant becomes less accurate after a deployment or prompt change. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
The targeted acceptance test is to evaluate a fixed representative prompt set across both versions. The design should first compare versioned evaluation scores against a prechange baseline and relevant user cohorts. A robust financial services division deployment also needs to require a stable regression test suite and deployment annotations; passing this check provides relevant evidence but does not certify every possible input or security risk.
Incorrect Answers
For token cost anomaly, it can be sensible to compare total tokens and spend before and after a traced optimization if overnight agent runs begin consuming far more output tokens than usual. Without the actual change to evaluate a fixed representative prompt set across both versions, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
The action to compare versioned evaluation scores against a prechange baseline and relevant user cohorts can contribute to model output drift as a implementation step; here, an assistant becomes less accurate after a deployment or prompt change. In the financial services division acceptance review, trying to compare versioned evaluation scores against a prechange baseline and relevant user cohorts does not test the exact model output drift condition where an assistant becomes less accurate after a deployment or prompt change; the relevant test is to evaluate a fixed representative prompt set across both versions.
In the financial services division acceptance review, trying to treat reduced complaint volume as definitive evidence of unchanged model quality does not test the exact model output drift condition where an assistant becomes less accurate after a deployment or prompt change; the relevant test is to evaluate a fixed representative prompt set across both versions. Acceptance should instead demonstrate that the service can evaluate a fixed representative prompt set across both versions, using the described failure as the test criterion.
Question 176
At a nonprofit organization, the security engineer is preparing a design meeting. The concrete requirement is: A knowledge assistant gives old policy answers after nightly ingestion. Which design action BEST addresses the stated need?
Explanation
Correct Answer
Choosing to inspect indexer status document freshness and failed ingestion counts directly targets the described search index health gap: a knowledge assistant gives old policy answers after nightly ingestion. As a lasting boundary, they also need to alert on ingestion lag and reject stale high-risk policy evidence, because an implementation by itself does not prove safe operation.
Incorrect Answers
The option to treat retrieved text as untrusted content and evaluate prompt-injection resistance would normally serve a different grounding attack monitoring requirement, where a retrieved page contains instructions attempting to override an agent system policy. The alternative may be useful in its own context, but it does not directly supply inspect indexer status document freshness and failed ingestion counts for the stated decision.
The action to modify a test source and verify its latest passage becomes retrievable can contribute to search index health as a targeted acceptance check step; here, a knowledge assistant gives old policy answers after nightly ingestion. Choosing to modify a test source and verify its latest passage becomes retrievable cannot directly replace the specified action to inspect indexer status document freshness and failed ingestion counts when a knowledge assistant gives old policy answers after nightly ingestion.
Choosing to assume an available search endpoint guarantees current indexed data cannot directly replace the specified action to inspect indexer status document freshness and failed ingestion counts when a knowledge assistant gives old policy answers after nightly ingestion. The actual implementation must inspect indexer status document freshness and failed ingestion counts, because the stated service condition is that a knowledge assistant gives old policy answers after nightly ingestion.
The action to alert on ingestion lag and reject stale high-risk policy evidence can contribute to search index health as a ongoing preventive rule step; here, a knowledge assistant gives old policy answers after nightly ingestion. Choosing to alert on ingestion lag and reject stale high-risk policy evidence cannot directly replace the specified action to inspect indexer status document freshness and failed ingestion counts when a knowledge assistant gives old policy answers after nightly ingestion.
Question 177
After a staged service rollout at a municipal services team, the team observes this failure: A knowledge assistant gives old policy answers after nightly ingestion. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
The faulty assumption is assume an available search endpoint guarantees current indexed data. A functioning alternative would be to inspect indexer status document freshness and failed ingestion counts, and the relevant validation would be to modify a test source and verify its latest passage becomes retrievable.
Incorrect Answers
The alternative to inspect indexer status document freshness and failed ingestion counts is a implementation step rather than the faulty assumption being diagnosed. This municipal services team failure-analysis alternative of inspect indexer status document freshness and failed ingestion counts does not diagnose the particular search index health defect where a knowledge assistant gives old policy answers after nightly ingestion; the keyed mistake is assume an available search endpoint guarantees current indexed data.
This municipal services team failure-analysis alternative of assume application authentication automatically restricts administrator permissions does not diagnose the particular search index health defect where a knowledge assistant gives old policy answers after nightly ingestion; the keyed mistake is assume an available search endpoint guarantees current indexed data. The alternative may be useful in its own context, but it does not directly supply assume an available search endpoint guarantees current indexed data for the stated decision.
The alternative to modify a test source and verify its latest passage becomes retrievable is a targeted acceptance check step rather than the faulty assumption being diagnosed. This municipal services team failure-analysis alternative of modify a test source and verify its latest passage becomes retrievable does not diagnose the particular search index health defect where a knowledge assistant gives old policy answers after nightly ingestion; the keyed mistake is assume an available search endpoint guarantees current indexed data.
Question 178
The payment services team is responding to a release acceptance review. A responsible-AI reviewer discovers this issue: A knowledge assistant gives old policy answers after nightly ingestion. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The appropriate remediation is to rerun failed indexer jobs and reconcile source versions, because the failure concerns search index health: a knowledge assistant gives old policy answers after nightly ingestion. A meaningful verification should modify a test source and verify its latest passage becomes retrievable, and the durable operating rule should alert on ingestion lag and reject stale high-risk policy evidence after the immediate fix.
Incorrect Answers
During payment services team remediation, the proposed step to assume an available search endpoint guarantees current indexed data is not the direct restoration action for search index health when a knowledge assistant gives old policy answers after nightly ingestion; the keyed repair is rerun failed indexer jobs and reconcile source versions. After remediation the team must modify a test source and verify its latest passage becomes retrievable; the action in this option does not itself establish that target recovery outcome.
The action to modify a test source and verify its latest passage becomes retrievable can contribute to search index health as a targeted acceptance check step; here, a knowledge assistant gives old policy answers after nightly ingestion. During payment services team remediation, the proposed step to modify a test source and verify its latest passage becomes retrievable is not the direct restoration action for search index health when a knowledge assistant gives old policy answers after nightly ingestion; the keyed repair is rerun failed indexer jobs and reconcile source versions.
This approach relates to token cost anomaly: overnight agent runs begin consuming far more output tokens than usual. If the team chose it instead, the targeted check to modify a test source and verify its latest passage becomes retrievable could still fail because the underlying required action is to rerun failed indexer jobs and reconcile source versions.
Question 179
A solution architect at a transport authority wants the following issue to be prevented in later deployments: A knowledge assistant gives old policy answers after nightly ingestion. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
The sustained prevention measure is to alert on ingestion lag and reject stale high-risk policy evidence. The implementation can still use the separate action to inspect indexer status document freshness and failed ingestion counts, but a guardrail constrains future changes. The transport authority can measure whether the rule works by arranging to modify a test source and verify its latest passage becomes retrievable.
Incorrect Answers
The action to tag every trace with deployment and prompt template identifiers addresses experiment reproducibility rather than the decisive search index health condition. The transport authority should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to alert on ingestion lag and reject stale high-risk policy evidence.
The action to rerun failed indexer jobs and reconcile source versions can contribute to search index health as a corrective remediation step; here, a knowledge assistant gives old policy answers after nightly ingestion. For prevention at the transport authority, the measure to rerun failed indexer jobs and reconcile source versions is weaker or aimed at another issue than the search index health guardrail to alert on ingestion lag and reject stale high-risk policy evidence needed when a knowledge assistant gives old policy answers after nightly ingestion.
For prevention at the transport authority, the measure to assume an available search endpoint guarantees current indexed data is weaker or aimed at another issue than the search index health guardrail to alert on ingestion lag and reject stale high-risk policy evidence needed when a knowledge assistant gives old policy answers after nightly ingestion. The lasting guardrail for this specific threat is to alert on ingestion lag and reject stale high-risk policy evidence, not simply to complete an adjacent service task.
Question 180
During a staged service rollout, a machine-learning developer must accept a change for search index health. The issue to be addressed is: A knowledge assistant gives old policy answers after nightly ingestion. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the large online retailer should modify a test source and verify its latest passage becomes retrievable. Operators still need the implementation to inspect indexer status document freshness and failed ingestion counts and the preventive requirement to alert on ingestion lag and reject stale high-risk policy evidence.
Incorrect Answers
In the large online retailer acceptance review, trying to assume an available search endpoint guarantees current indexed data does not test the exact search index health condition where a knowledge assistant gives old policy answers after nightly ingestion; the relevant test is to modify a test source and verify its latest passage becomes retrievable. Acceptance should instead demonstrate that the service can modify a test source and verify its latest passage becomes retrievable, using the described failure as the test criterion.
For private data telemetry, it can be sensible to submit a known sensitive test string and inspect who can retrieve telemetry if model traces risk capturing customer health details in plaintext. Without the actual change to modify a test source and verify its latest passage becomes retrievable, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
The action to inspect indexer status document freshness and failed ingestion counts can contribute to search index health as a implementation step; here, a knowledge assistant gives old policy answers after nightly ingestion. In the large online retailer acceptance review, trying to inspect indexer status document freshness and failed ingestion counts does not test the exact search index health condition where a knowledge assistant gives old policy answers after nightly ingestion; the relevant test is to modify a test source and verify its latest passage becomes retrievable.
Developing AI Apps and Agents on Azure (AI-103) Practice Tests
- Microsoft AI-103 Practice Test 16 - 10 Questions (Last Reviewed 06-Nov-2026)
- Microsoft AI-103 Practice Test 17 - 10 Questions (Last Reviewed 05-Nov-2026)
- Microsoft AI-103 Practice Test 18 - 10 Questions (Last Reviewed 04-Nov-2026)
- Microsoft AI-103 Practice Test 19 - 10 Questions (Last Reviewed 03-Nov-2026)
- Microsoft AI-103 Practice Test 20 - 10 Questions (Last Reviewed 02-Nov-2026)