Microsoft AI-103 Practice Test 16
Question 150
A search index returns archived policies even after records are replaced with a new signed version. What index update strategy is required?
- Retain all versions with no effective-date or active-state filtering.
- Rebuild only the vector embeddings and ignore the source version mapping.
- Change search-result ranking by document length rather than state.
- Use stable document IDs with change detection and retire superseded chunks from search.
Explanation
Correct Answer
D:Incremental ingestion must update and remove old versions, not leave duplicates indefinitely searchable.
Incorrect Answers
A:Obsolete records can remain retrievable and contaminate answers even after a new policy is signed.
B:Recomputing embedding vectors does not remove outdated policy versions or change which source revision is effective. The indexer must reconcile original document IDs and mark superseded chunks unavailable for current retrieval.
C:Length cannot distinguish an approved current policy from an archived revision.
Question 151
A retrieval chatbot cites correct policies but its answers are too long for mobile devices. What dimension should evaluation include?
- Evaluate only whether citations are technically valid, ignoring whether the mobile screen can display the answer usefully.
- Use citation-groundedness as the only release metric and disregard mobile readability or excessive answer length.
- Assess output length and usability against a mobile-oriented answer rubric
- Measure both false-negative safety escapes and false-positive refusals
Explanation
Correct Answer
C:Response usefulness includes meeting interface constraints as well as factual correctness. This directly supports the need to test answer concision for a specified interface, unlike a mechanism primarily intended to control variance in model-based evaluation.
Incorrect Answers
A:A well-grounded answer can still fail on the intended device. Assess length, scannability and task usability against representative mobile interactions.
B:A correct, source-backed answer can still be unusable on small screens. Evaluate word budget, scanability, and mobile task completion separately from factual groundedness.
D:Moderation improvements require balancing actual harms against excessive blocking. Its primary purpose is to evaluate safety overblocking and missed harms; the stated constraint instead demands a way to test answer concision for a specified interface.
Question 152
Two Foundry environments serve different business units. A developer requests access only to development assets and not production. How should access be given?
- Assign the appropriate Foundry/Azure role at the least-privileged project or resource scope.
- Give the developer subscription Owner on both development and production.
- Share an administrator's deployment key between both environments.
- Grant production project roles because the developer has access to its model catalog.
Explanation
Correct Answer
A:Role scoping limits the developer's access to the intended environment while preserving separate production administration and audit trails.
Incorrect Answers
B:Subscription-wide Owner would expand privileges into the protected production environment.
C:Shared privileged secrets defeat user accountability and separation between environments.
D:Catalog visibility does not justify production resource or inference permissions.
Question 153
A real-estate assistant describes a room with 'a west-facing window' from a photo lacking orientation metadata. What should it avoid?
- Assign compass directions to photographed windows by assuming the image's left side always corresponds to west.
- Infer the room's compass orientation from the brightness of the window, even though the image has no trustworthy directional metadata.
- Exclude unsupported orientation claims unless grounded in trusted context
- Request evidence-grounded side-by-side comparison of marked features
Explanation
Correct Answer
C:A visual image often cannot establish compass directions without additional evidence. The decisive difference is whether the design can prevent unsupported visual-context inferences; the competing function serves to describe several images with source separation.
Incorrect Answers
A:Image orientation is not inferable from pixel-left versus geographic west without trusted camera or location metadata. Report only visible window placement and label compass orientation unknown if unsupported.
B:Visible daylight does not establish west-facing orientation. The description should limit itself to observable features unless external orientation evidence exists.
D:A comparative visual task should identify observed differences rather than assert unsupported product identity. That is relevant to compare two visually similar components carefully, but the present scenario requires an ability to prevent unsupported visual-context inferences.
Question 154
A customer chat tool detects abusive personal threats and must route urgent cases to trained staff. What workflow is appropriate?
- Classify an explicit personal threat as ordinary negative sentiment and respond with only a generic apology
- Trigger a policy-based safety escalation for high-severity threat classifications
- Publish the threatening message as an ordinary low-priority chat summary without triggering the organization’s trained escalation process.
- Evaluate contextual safety classifications rather than using keywords alone
Explanation
Correct Answer
B:High-risk text events require defined handling rather than silent standard chatbot continuation. The decisive difference is whether the design can escalate threatening content for human review; the competing function serves to prevent exposure of credentials in model output.
Incorrect Answers
A:Threat detection calls for a severity-aware safety response, not merely a polite sentiment-handling template. A trained escalation route is needed for high-risk cases.
C:High-severity threat detection requires policy-defined escalation and accountable human handling; routine summarization alone does not address urgent risk.
D:Unsafe intent depends on context; simple word matches produce overblocking. That is relevant to reduce false-positive text moderation, but the present scenario requires an ability to escalate threatening content for human review.
Question 155
An application wants to evaluate the first five evidence results for each question. What metric fits?
- Measure precision at five against labeled relevant passages
- Select five retrieval results in their raw index order without evaluating relevance to the query
- Use the overall number of returned documents as the performance metric without knowing which of the top five are relevant.
- Use semantic retrieval or vector search over appropriately embedded passages
Explanation
Correct Answer
A:Precision at k directly measures how often top results meet user information needs. Here the relevant capability must let the team score accuracy of top search results. An alternative targeting place precise evidence above generic matches solves another problem.
Incorrect Answers
B:Top-k count controls quantity, not evidential usefulness. Results must be scored or judged against the query and evaluated for answer support.
C:Result count is not retrieval precision. Precision at five requires labeled relevance judgments for the first five ranked evidence results.
D:Semantic similarity can surface conceptually related text despite vocabulary mismatch. Here the needed outcome is to score accuracy of top search results; an option designed to retrieve paraphrased source passages would not directly provide it.
Question 156
An agent completes user requests but sometimes executes unapproved API operations. Which assessment scope is missing?
- Evaluate only the final natural-language response and ignore the actions invoked along the agent's tool path.
- Evaluate only the quality of the final response and ignore the intermediate API operations the agent actually executed.
- Evaluate the agent tool trajectory and policy compliance, not just final text
- Measure task fulfillment and relevance against representative user goals
Explanation
Correct Answer
C:A successful final answer does not justify unauthorized intermediate actions. The decisive difference is whether the design can evaluate tool-call behavior for safety; the competing function serves to detect unsupported factual claims.
Incorrect Answers
A:A good-looking final message can follow an unauthorized operation. Validate the full tool trajectory, permissions, parameters and side effects, including blocked or redirected attempts.
B:A plausible final answer can hide unauthorized tool actions. Review the tool trajectory against allowlists, approvals and permitted side effects.
D:Keyword coverage alone does not prove the answer resolves the user's problem. That is relevant to evaluate real task success, but the present scenario requires an ability to evaluate tool-call behavior for safety.
Question 157
An agent continues running after a compute failover but loses every grounding document when its only storage zone fails. What architecture improvement is missing?
- Create a tagging policy initiative but leave all grounding documents in one unreplicated availability-zone volume.
- Increase the controller's compute failover capacity while keeping all grounding documents in the same single-failure-zone store.
- Add more application controllers but retain the only grounding-document copy in an unreplicated single-zone volume.
- Plan resilient grounding-data storage and test restoration for the required failure scope
Explanation
Correct Answer
D:Redundant controllers cannot answer grounded questions if their entire document store disappears; durability, recovery and data-location requirements must be designed for the data dependency too. In this scenario the team must maintain usable grounding data after a storage failure. A feature used to enforce declarative resource governance requirements does not substitute for it.
Incorrect Answers
A:A governance initiative does not restore data lost with a storage-zone failure. Replicate or protect grounding data to meet the resilience objective, and verify recovery procedures.
B:Compute redundancy cannot restore inaccessible knowledge data. Replicate or back up grounding storage for the required failure scope and test recovery.
C:Compute redundancy cannot restore missing knowledge content. Protect the grounding data with an appropriate resilient storage design and test that the application can retrieve it after regional or zone disruptions.
Question 158
An app answers questions about a screenshot containing a chart, but ignores axis labels. What should its visual evaluation check?
- Evaluate the chart screenshot using only object labels and ignore its axis direction, numeric scales, units and legend.
- Verify chart-reading accuracy including axes, units and legends
- Describe chart colors and overall page theme while ignoring the axis values, units, tick marks and legend.
- Express uncertainty when the image lacks sufficient resolution to support recognition
Explanation
Correct Answer
B:Visual numeric interpretation requires more than recognizing colored bars or lines. In this scenario the team must interpret chart values with labels and units. A feature used to prevent unsupported visual-context inferences does not substitute for it.
Incorrect Answers
A:A model may recognize bars but misstate the result if it ignores graph semantics. Evaluation must verify axes, units, legends and numeric relationships.
C:Chart understanding depends on extracting scales and labels, not just visual styling. Check each plotted value against the appropriate axis and legend before explaining trends.
D:Visual evidence limits must govern claims; hallucinating obscured text can cause operational errors. This helps when a team needs to avoid inventing unreadable visual details; it does not establish how to interpret chart values with labels and units.
Question 159
A model labels a customer complaint safe but outputs a secret API key quoted from a log. What separate filter is needed?
- Apply secret-detection and output scanning independent of broad safety categories
- Assume a general toxicity filter will detect every exposed API token in a response
- Rely only on a generic harmful-language classifier and publish the quoted access token because the surrounding complaint sounds harmless.
- Trigger a policy-based safety escalation for high-severity threat classifications
Explanation
Correct Answer
A:A response can be nonviolent yet still leak credentials; secret-specific scanning closes that gap. In this scenario the team must prevent exposure of credentials in model output. A feature used to reduce false-positive text moderation does not substitute for it.
Incorrect Answers
B:An otherwise polite output can still reveal a credential. Output-side secret scanning is a separate check from category-level safety or abuse detection.
C:An otherwise safe complaint can still disclose credentials. Apply independent secret detection and output redaction before sharing model text.
D:High-risk text events require defined handling rather than silent standard chatbot continuation. This helps when a team needs to escalate threatening content for human review; it does not establish how to prevent exposure of credentials in model output.
Question 160
A semantic search prototype uses only one paraphrased query as its benchmark. What evaluation improvement matters?
- Declare search quality satisfactory based on a single paraphrase and no representative negatives
- Benchmark the retriever on one hand-picked query and report success without testing short queries, jargon, synonyms or paraphrases.
- Measure precision at five against labeled relevant passages
- Build a diverse labeled query set covering short and long phrasing
Explanation
Correct Answer
D:Relevance assessment needs representative query diversity before claiming robust semantic quality. This directly supports the need to evaluate search across different query phrasings, unlike a mechanism primarily intended to retrieve paraphrased source passages.
Incorrect Answers
A:One successful query cannot measure robustness to wording variation or false matches. A held-out set of representative queries and hard negatives is needed.
B:A single favorable phrasing cannot establish retrieval robustness. Build a diverse labeled query set representing the real language variations users will submit.
C:Precision at k directly measures how often top results meet user information needs. Here the needed outcome is to evaluate search across different query phrasings; an option designed to score accuracy of top search results would not directly provide it.