Microsoft AI-103 Practice Test 14
Question 130
A marketing repository includes image captions but no ability to find pictures by semantic visual similarity. What additional field helps?
- Index image file names as text and assume that alone supports semantic visual similarity
- Compute supported image embeddings and index them for vector search
- Index image filenames and their human-written captions only, without computing any embedding from the actual pixels.
- Preserve modality, asset and temporal relationships across indexed segments
Explanation
Correct Answer
B:Visual embeddings can represent image content for semantic retrieval beyond caption terms. The decisive difference is whether the design can search image collections by visual similarity; the competing function serves to link scene results to original footage.
Incorrect Answers
A:File names rarely encode visual content reliably. Images need an appropriate visual representation, such as supported embeddings, for semantic visual search.
C:Text metadata cannot reliably support visual similarity between images with different captions. Visual embeddings are needed for nearest-neighbor image search.
D:Typed links prevent text from one source segment being misattributed to another. That is relevant to maintain multimodal source associations, but the present scenario requires an ability to search image collections by visual similarity.
Question 131
A candidate model gives polished summaries but invents document-specific regulations. Which evaluator detects this?
- Score only reading ease and elegant prose while treating every unsupported regulatory statement as acceptable.
- Calibrate the rubric and compare judgments with human-reviewed anchor examples
- Score groundedness against source references, not only fluency
- Measure task fulfillment and relevance against representative user goals
Explanation
Correct Answer
C:A polished response can be factually wrong when it cites invented regulatory obligations. Compare each assertion with the retrieved source passages and score evidential support and citation accuracy; presentation fluency alone cannot establish groundedness.
Incorrect Answers
A:Polished language can still invent legal obligations. A groundedness evaluator must compare individual regulatory claims to the actual cited documents.
B:Repeatability and anchored scoring improve confidence in automatic evaluation. That is relevant to control variance in model-based evaluation, but the present scenario requires an ability to detect unsupported factual claims.
D:Keyword coverage alone does not prove the answer resolves the user's problem. The function serves to evaluate real task success, whereas the actual requirement is to detect unsupported factual claims.
Question 132
An agent provides a document citation but the document's ACL would deny this user access. What should change?
- Enforce document-level authorization before evidence enters agent context
- Allow the model to read restricted passages before applying the user's access rules solely to the final list of citations.
- Validate an action tool's JSON arguments but omit document-level ACL filtering on retrieved citations.
- Limit conversation-memory length but permit cited confidential passages to enter context without ACL checks.
Explanation
Correct Answer
A:Retrieval security must use the requester's entitlements; a citation should not reveal restricted text. The practical requirement is to filter retrieved knowledge by the user's permission; a solution built to ground answers in governed current documents would not meet that same need.
Incorrect Answers
B:Late citation filtering does not stop restricted evidence entering model context. Authorization must be enforced before retrieval results are exposed to the agent.
C:A valid action schema does not authorize a user's access to a confidential document. Enforce document permissions before the passage enters agent context.
D:Bounded session history doesn't authorize document access. Retrieval must enforce source permissions before any prohibited passage can be cited or used.
Question 133
A museum app receives an image and must answer 'How many chairs are visible?' based on actual pixels. What input mode is needed?
- Send the image and question to a multimodal visual-understanding model
- Generate a prose caption from the image filename and infer chair count without examining the pixel data.
- Answer chair counts from the photo's file name and tags without including the uploaded picture in the visual model request.
- Prompt for a concise factual image caption centered on salient visible content
Explanation
Correct Answer
A:The model must inspect the visual scene rather than answer from the user's words alone. This resolves the specific need to answer an image-grounded visual count, while a measure that helps describe several images with source separation addresses a different requirement.
Incorrect Answers
B:Object-counting requires visual evidence in the submitted picture. Filename or caption metadata cannot establish how many chairs are visible or distinguish occluded objects.
C:Metadata may not list all visible objects. The count must be grounded in the image pixels supplied to a compatible multimodal vision model.
D:A short factual caption should describe what is visible without unrelated speculation. The function serves to produce compact visually grounded captions, whereas the actual requirement is to answer an image-grounded visual count.
Question 134
An enterprise classifier repeatedly calls sarcasm positive because it sees praise words. What improvement helps?
- Use contextual examples and evaluate sarcasm-specific errors
- Count favorable adjectives and skip evaluation on sarcastic examples
- Treat the presence of positive adjectives as a conclusive positive classification, without testing sarcastic wording or context.
- Classify sentiment polarity from the customer's feedback
Explanation
Correct Answer
A:Literal polarity words can contradict contextual meaning, so domain testing is necessary. The practical requirement is to recognize sentiment polarity in sarcastic context; a solution built to identify mixed sentiment about separate product aspects would not meet that same need.
Incorrect Answers
B:Sarcasm may pair praise words with a negative implied judgment. Literal keyword counting without targeted error evaluation will reproduce the identified failure.
C:Literal praise words can express criticism through sarcasm. Train and evaluate on context-sensitive counterexamples instead of relying on keyword polarity.
D:Sentiment captures evaluative polarity rather than merely detecting subject keywords. That capability is appropriate for measure positive, negative or neutral opinion, not this separate task of finding how to recognize sentiment polarity in sarcastic context.
Question 135
A mixed-media ingestion service merges thumbnail captions with unrelated transcript segments. How should records be structured?
- Flatten all captions and transcripts into one undifferentiated text document before retrieval
- Join transcript and thumbnail records solely by their common file extension, ignoring the scene timeline and media asset identifiers.
- Preserve modality, asset and temporal relationships across indexed segments
- Compute supported image embeddings and index them for vector search
Explanation
Correct Answer
C:Typed links prevent text from one source segment being misattributed to another. In this scenario the team must maintain multimodal source associations. A feature used to create searchable video-scene metadata does not substitute for it.
Incorrect Answers
A:Flattening removes which modality and moment produced each claim. An asset- and time-aware structure avoids attaching a caption to an unrelated audio segment.
B:File extensions do not establish which frames align with spoken segments. Preserve source asset, modality and precise time relationships in the index.
D:Visual embeddings can represent image content for semantic retrieval beyond caption terms. This helps when a team needs to search image collections by visual similarity; it does not establish how to maintain multimodal source associations.
Question 136
Marketing says a model response is good because it mentions every keyword, yet users cannot act on it. What should be scored?
- Mark an answer successful whenever it reproduces the requested vocabulary, without testing whether the user can complete their task.
- Evaluate the agent tool trajectory and policy compliance, not just final text
- Score only whether the response includes the same keywords as the request, regardless of whether the task is completed.
- Measure task fulfillment and relevance against representative user goals
Explanation
Correct Answer
D:Keyword coverage alone does not prove the answer resolves the user's problem. The practical requirement is to evaluate real task success; a solution built to control variance in model-based evaluation would not meet that same need.
Incorrect Answers
A:Keyword coverage does not prove that the response solves the user's problem. Evaluate actual task fulfillment against representative user scenarios.
B:A successful final answer does not justify unauthorized intermediate actions. This helps when a team needs to evaluate tool-call behavior for safety; it does not establish how to evaluate real task success.
C:Keyword overlap is a weak proxy for actionability. Evaluate whether representative users can accomplish their actual task and whether the answer addresses the expressed intent with relevant, usable details.
Question 137
An agent starts repeating obsolete instructions from its full conversation transcript. What context practice reduces this risk?
- Replay all historical conversation text at full length, including superseded instructions, whenever the agent makes a new decision.
- Expose a transcript-write action and keep replaying every obsolete instruction rather than pruning stale context.
- Summarize or prune stale conversation state while retaining verified task constraints
- Enforce document-level authorization before evidence enters agent context
Explanation
Correct Answer
C:Selective context reduces token bloat and avoids treating old statements as current authoritative instructions. Here the relevant capability must let the team maintain compact current context. An alternative targeting isolate memory between users solves another problem.
Incorrect Answers
A:Unbounded replay gives obsolete instructions continuing influence. Summarize or prune stale context while retaining verified current constraints and approvals.
B:A write tool cannot distinguish obsolete conversational instructions on its own. Summarize or prune context while retaining verified current task constraints.
D:Retrieval security must use the requester's entitlements; a citation should not reveal restricted text. Here the needed outcome is to maintain compact current context; an option designed to filter retrieved knowledge by the user's permission would not directly provide it.
Question 138
An accessibility app needs one short caption describing the primary scene of a photograph. What output contract is appropriate?
- Return a speculative description of the photographer's intentions and the people shown, instead of a short statement of visible scene content.
- Verify chart-reading accuracy including axes, units and legends
- Express uncertainty when the image lacks sufficient resolution to support recognition
- Prompt for a concise factual image caption centered on salient visible content
Explanation
Correct Answer
D:A short factual caption should describe what is visible without unrelated speculation. The practical requirement is to produce compact visually grounded captions; a solution built to prevent unsupported visual-context inferences would not meet that same need.
Incorrect Answers
A:Accessibility captions should describe salient visible facts concisely. Inferred motivations or unsupported personal details are unnecessary and can mislead listeners.
B:Visual numeric interpretation requires more than recognizing colored bars or lines. This helps when a team needs to interpret chart values with labels and units; it does not establish how to produce compact visually grounded captions.
C:Visual evidence limits must govern claims; hallucinating obscured text can cause operational errors. The function serves to avoid inventing unreadable visual details, whereas the actual requirement is to produce compact visually grounded captions.
Question 139
A support transcript contains social security numbers that must not be exposed in summaries. What processing step fits?
- Post the transcript in full to a shared channel and ask readers to ignore any Social Security numbers they see.
- Apply secret-detection and output scanning independent of broad safety categories
- Detect and redact protected personal identifiers before sharing the output
- Evaluate contextual safety classifications rather than using keywords alone
Explanation
Correct Answer
C:Sensitive-data recognition and masking reduce privacy leakage from otherwise useful summaries. Here the relevant capability must let the team protect sensitive identifiers in text outputs. An alternative targeting escalate threatening content for human review solves another problem.
Incorrect Answers
A:An instruction to readers does not remove sensitive identifiers. Detect and redact protected values before disclosure, with access controls for any restricted original.
B:A response can be nonviolent yet still leak credentials; secret-specific scanning closes that gap. The function serves to prevent exposure of credentials in model output, whereas the actual requirement is to protect sensitive identifiers in text outputs.
D:Unsafe intent depends on context; simple word matches produce overblocking. Here the needed outcome is to protect sensitive identifiers in text outputs; an option designed to reduce false-positive text moderation would not directly provide it.
Question 140
A user asks a natural-language policy question using words not present in the relevant source. What retrieval mechanism helps?
- Require every user query to repeat the source's exact technical vocabulary before permitting a result to match.
- Build a diverse labeled query set covering short and long phrasing
- Apply semantic reranking to top retrieval candidates
- Use semantic retrieval or vector search over appropriately embedded passages
Explanation
Correct Answer
D:Semantic similarity can surface conceptually related text despite vocabulary mismatch. This resolves the specific need to retrieve paraphrased source passages, while a measure that helps score accuracy of top search results addresses a different requirement.
Incorrect Answers
A:A natural-language paraphrase may express the same meaning with different words. Semantic or vector retrieval can bridge that wording gap.
B:Relevance assessment needs representative query diversity before claiming robust semantic quality. That is relevant to evaluate search across different query phrasings, but the present scenario requires an ability to retrieve paraphrased source passages.
C:Reranking can improve the order of already retrieved passages based on query relevance. This helps when a team needs to place precise evidence above generic matches; it does not establish how to retrieve paraphrased source passages.