Microsoft AI-103 Practice Test 36
Question 350
A tool response contains document titles but no source IDs, and the model invents citations. What should be provided?
- Return stable source identifiers and evidence spans with search results
- Cite a plausible document title without a stable source identifier or exact retrieved span
- Set result limits and relevance thresholds appropriate to the context budget
- Expose an authorized search tool with a bounded query interface to the agent
Explanation
Correct Answer
A:Grounded citations require identifiers that link generated claims to actual retrieved passages. The decisive difference is whether the design can supply verifiable citation handles to an agent; the competing function serves to enforce search authorization for agent tools.
Incorrect Answers
B:A title alone cannot substantiate a specific statement or support an audit trail. The tool response needs verifiable source identity and location.
C:Bounded search responses preserve useful evidence while controlling tokens and noise. Its primary purpose is to limit and rank agent search-tool results; the stated constraint instead demands a way to supply verifiable citation handles to an agent.
D:Tool integration lets the agent retrieve relevant knowledge when needed without embedding entire corpora. That is relevant to enable an agent to query indexed evidence, but the present scenario requires an ability to supply verifiable citation handles to an agent.
Question 351
A multi-agent workflow includes dependent investigations, concurrent edits to a shared file, and occasional failed specialists. Which THREE orchestration controls address those issues? Choose THREE.
- Track each task as pending, complete or failed and surface partial results accurately
- Dispatch only the capabilities needed by the request
- Require source references, scope and confidence in each specialist deliverable
- Run dependent investigations concurrently, permit unsynchronized edits to a shared file and suppress failed-agent states.
- Give one owner write authority or serialize merges through the coordinator
- Model prerequisite dependencies and parallelize only independent tasks
Explanation
Correct Answers
A:Coordinators should not conceal a failed workstream or fabricate missing evidence. Dependency sequencing, shared-write control and task lifecycle status address distinct coordination hazards.
E:Concurrent stateful writes need ownership and a merge discipline to prevent lost changes.
F:Legal jurisdiction is required input to privacy assessment while pricing has no such dependency.
Incorrect Answers
B:Selective orchestration avoids unnecessary model calls, latency and possible conflicting outputs. Its suitable role is to avoid unnecessary agent fan-out; it would leave the specific need to schedule dependencies, serialize writes and track failures unresolved.
C:Traceable outputs let the coordinating agent verify and cite claims instead of merging unsupported summaries. Here, however, the stated objective is to schedule dependencies, serialize writes and track failures; this option focuses on make specialist results auditable.
D:The combined workflow needs prerequisites, serialized shared writes and explicit partial-failure reporting. Removing all three controls makes results unreliable.
Question 352
A production safety dashboard lacks evidence showing why the assistant refused a request. What observability should be added?
- Log all full customer prompts and raw secrets without access controls.
- Track only whether the HTTP request succeeded or failed.
- Disable the safety filter to avoid refusals rather than measure them.
- Capture privacy-safe filter decisions, policy version, evaluation labels and trace correlation IDs.
Explanation
Correct Answer
D:Safety telemetry lets engineers explain and investigate refusals without storing unnecessary raw sensitive prompts or weakening the safety policy.
Incorrect Answers
A:Unrestricted raw logging exposes sensitive data and is unnecessary for diagnosing safety refusals.
B:Transport success does not explain which safety policy or filter produced a refusal.
C:Suppressing protection prevents the requested auditability and can expose unsafe outputs.
Question 353
A model flags a person near a restricted doorway but the app needs coordinates for highlighting the individual. What output does it need?
- Produce a free-form caption such as 'person near door' without returning any image coordinates or region geometry.
- Request object-region localization with coordinates for the source frame
- Segment the video into time-stamped intervals with searchable metadata
- Associate extracted frame features with source video IDs and segment offsets
Explanation
Correct Answer
B:Region or bounding-box outputs identify the location, not just presence, of an object. Here the relevant capability must let the team locate a detected subject in a frame. An alternative targeting preserve object identity across video frames solves another problem.
Incorrect Answers
A:A textual caption confirms the observed subject but cannot position a highlight. Localization requires bounding coordinates or another supported region representation for the source frame.
C:Temporal chunking lets analysis and search refer to specific portions rather than treating video as one blob. That capability is appropriate for find an event in a large video by timestamp, not this separate task of finding how to locate a detected subject in a frame.
D:Provenance links allow retrieved frame evidence to be located in the original recording. Here the needed outcome is to locate a detected subject in a frame; an option designed to map search matches back to source video would not directly provide it.
Question 354
A custom recognition model performs well in a quiet office but poorly on industrial machines. What data change helps?
- Include representative background noise and device conditions in evaluation and adaptation
- Declare production readiness from quiet-room recognition scores alone
- Evaluate custom speech recognition with domain-specific labeled audio
- Verify speaker consent, rights and permitted synthetic-voice usage
Explanation
Correct Answer
A:Acoustic conditions strongly affect recognition, so realistic field recordings matter. The practical requirement is to validate recognition under real environmental noise; a solution built to evaluate speech recognition on unseen audio would not meet that same need.
Incorrect Answers
B:A model evaluated only on studio audio may fail on industrial noise and channel artifacts. Test on representative field recordings.
C:Custom speech can improve recognition for specialized vocabulary when adequate training data exists. The function serves to adapt ASR to industrial terminology, whereas the actual requirement is to validate recognition under real environmental noise.
D:Voice customization should not impersonate or clone real speakers without authorized rights. That capability is appropriate for obtain consent before creating a custom voice, not this separate task of finding how to validate recognition under real environmental noise.
Question 355
A user asks a broad question, and the agent search tool returns hundreds of irrelevant hits. What tool contract helps?
- Send hundreds of raw search matches directly to the model without ranking or context limits
- Apply permission and tenant filters before returning search results to the model
- Set result limits and relevance thresholds appropriate to the context budget
- Return stable source identifiers and evidence spans with search results
Explanation
Correct Answer
C:Bounded search responses preserve useful evidence while controlling tokens and noise. In this scenario the team must limit and rank agent search-tool results. A feature used to enable an agent to query indexed evidence does not substitute for it.
Incorrect Answers
A:Unfiltered result floods can bury decisive passages and exceed context budgets. Apply relevance ranking and bounded retrieval.
B:RAG tools must honor caller identity and document ACLs rather than relying on the model to self-filter. Its primary purpose is to enforce search authorization for agent tools; the stated constraint instead demands a way to limit and rank agent search-tool results.
D:Grounded citations require identifiers that link generated claims to actual retrieved passages. This helps when a team needs to supply verifiable citation handles to an agent; it does not establish how to limit and rank agent search-tool results.
Question 356
An agent loops on the same failed API call for 30 minutes. What external controller rule should stop it?
- Stop side effects while retaining governed logs and trace evidence
- Enlarge the agent's tenant permission scope and assume wider access stops its repeated failed API calls.
- Apply write-call idempotency keys while allowing an agent to repeat the same failing read operation indefinitely without a call budget.
- Enforce an iteration/time budget with no-progress detection and explicit failure
Explanation
Correct Answer
D:Bounded tool execution prevents runaway use even when the agent cannot self-recognize a loop. The decisive difference is whether the design can terminate runaway autonomous actions; the competing function serves to halt unsafe actions without destroying evidence.
Incorrect Answers
A:Safe termination should preserve information necessary to diagnose and review the violation. That capability is appropriate for halt unsafe actions without destroying evidence, not this separate task of finding how to terminate runaway autonomous actions.
B:Authorization boundaries do not impose iteration or retry limits. Stop the loop using an external call budget, repeated-error detection and a terminal failure state.
C:Idempotency reduces duplicate side effects on retried writes but does not stop runaway retry loops. The controller must impose time, iteration and no-progress limits even for safe read operations.
Question 357
An AI service can access its model deployment but fails to read blob-grounding content. What separate permission should be granted?
- Assign a narrowly scoped Storage Blob Data Reader role to its workload identity
- Reduce the principal's permissions to model-inference roles at the smallest applicable scope
- Disable the active service principal and remove its existing role assignments instead of granting scoped access to the required blob container.
- Use workload identity federation for the build pipeline
Explanation
Correct Answer
A:Model invocation permission is distinct from reading blobs through the storage data plane. In this scenario the team must permit read-only blob data retrieval. A feature used to enforce least privilege for model inference does not substitute for it.
Incorrect Answers
B:Broad Owner grants are unrelated to the specific model call and unnecessarily enable resource changes. Its primary purpose is to enforce least privilege for model inference; the stated constraint instead demands a way to permit read-only blob data retrieval.
C:The workload identity is still needed for serving the assistant. Deleting its grants would remove functioning model access and not provide the missing Blob data-plane read permission.
D:Federated tokens allow trusted CI identities to acquire Azure access without static client secrets. This helps when a team needs to authenticate automated deployments without client secrets; it does not establish how to permit read-only blob data retrieval.
Question 358
A video analysis app confuses two similar moving objects between frames. What enhancement is relevant?
- Independently label each video frame and discard cross-frame identity information before comparing movement.
- Choose temporal sampling based on event duration and risk of missing brief actions
- Compare temporally sampled frames while preserving their order
- Apply object tracking with temporal identity association
Explanation
Correct Answer
D:Tracking maintains correspondence across frames rather than making unrelated per-frame detections. This directly supports the need to preserve object identity across video frames, unlike a mechanism primarily intended to locate a detected subject in a frame.
Incorrect Answers
A:Standalone detections cannot reliably establish which similar object in the next frame corresponds to the first one. Temporal identity association is the missing tracking step.
B:Sparse sampling can miss transient events, so interval selection must reflect task demands. That capability is appropriate for balance frame processing cost against missed events, not this separate task of finding how to preserve object identity across video frames.
C:Visual reasoning about motion requires correct temporal sequence, not unordered snapshots. Here the needed outcome is to preserve object identity across video frames; an option designed to compare observed visual changes through time would not directly provide it.
Question 359
A support agent must infer whether a caller is asking a question or issuing an instruction, using both transcription and audio cues. What model setup fits?
- Classify intent from a text transcript alone while explicitly discarding the caller's audio and interaction context required by the task.
- Keep speaker-labeled timestamps aligned with recognized utterances
- Provide time-coded transcript excerpts and uncertainty flags
- Use an audio-capable multimodal reasoning workflow with explicit task criteria
Explanation
Correct Answer
D:Voice intent may depend on acoustic as well as lexical evidence; the model should be tested accordingly. Here the relevant capability must let the team interpret spoken intent using audio context. An alternative targeting avoid overconfident emotion attribution from audio solves another problem.
Incorrect Answers
A:A transcript may omit prosody and other audio cues relevant to disambiguating question versus instruction. Use the supported audio-capable workflow with defined criteria and avoid unsupported emotional inferences.
B:Time and speaker grounding makes later claims traceable to specific audio segments. The function serves to locate statements by speaker and time, whereas the actual requirement is to interpret spoken intent using audio context.
C:Grounded transcription references let reviewers revisit contested acoustic evidence. That capability is appropriate for trace audio reasoning back to evidence, not this separate task of finding how to interpret spoken intent using audio context.
Question 360
An invoice has line items in a table and totals beneath it. Which extraction representation helps accounting automation?
- Flatten all invoice cells into one string and drop the line-item row-to-amount associations.
- Use layout-aware extraction with table rows, columns and numeric fields
- Recover reading order based on detected document layout
- Use document analysis capable of representing detected layout elements and regions
Explanation
Correct Answer
B:Flat OCR text cannot reliably preserve which prices belong to which line items. This resolves the specific need to extract invoice table layout and numeric values, while a measure that helps locate extracted document fields in the source addresses a different requirement.
Incorrect Answers
A:Unstructured text can mix quantities, unit prices and totals. Layout-aware table extraction preserves column and row relationships needed for accurate invoice processing.
C:Layout-aware reading order is necessary for coherent extracted sentences across columns. This helps when a team needs to preserve multi-column reading sequence; it does not establish how to extract invoice table layout and numeric values.
D:Non-paragraph marks can be business evidence and need structured representation where supported. The function serves to capture non-textual form elements, whereas the actual requirement is to extract invoice table layout and numeric values.