Microsoft AI-103 Practice Test 35
Question 340
An agent retrieval tool returns confidential snippets from unrelated departments. What control should be added?
- Copy the retrieved HR policy passage into every team response without checking the caller's authorization
- 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
B:RAG tools must honor caller identity and document ACLs rather than relying on the model to self-filter. This directly supports the need to enforce search authorization for agent tools, unlike a mechanism primarily intended to enable an agent to query indexed evidence.
Incorrect Answers
A:A retrieved text match is not a disclosure authorization. Source access restrictions must be enforced before the agent returns a confidential passage.
C:Bounded search responses preserve useful evidence while controlling tokens and noise. Here the needed outcome is to enforce search authorization for agent tools; an option designed to limit and rank agent search-tool results would not directly provide it.
D:Grounded citations require identifiers that link generated claims to actual retrieved passages. Its primary purpose is to supply verifiable citation handles to an agent; the stated constraint instead demands a way to enforce search authorization for agent tools.
Question 341
A review team debates two agents' inconsistent regulatory interpretations. Who should decide which conclusion to publish?
- Dispatch only the capabilities needed by the request
- Use evidence-based conflict adjudication by the coordinator or authorized reviewer
- Automatically publish the interpretation from whichever regulatory specialist finishes first, without reconciliation.
- Track each task as pending, complete or failed and surface partial results accurately
Explanation
Correct Answer
B:Specialist outputs are hypotheses requiring primary-source reconciliation when they disagree. Here the relevant capability must let the team reconcile contradictory specialist claims. An alternative targeting avoid unnecessary agent fan-out solves another problem.
Incorrect Answers
A:Selective orchestration avoids unnecessary model calls, latency and possible conflicting outputs. This helps when a team needs to avoid unnecessary agent fan-out; it does not establish how to reconcile contradictory specialist claims.
C:Completion order is not an evidence or authority ranking. A coordinator or accountable reviewer should adjudicate conflicting claims against the applicable source rules before publication.
D:Coordinators should not conceal a failed workstream or fabricate missing evidence. Here the needed outcome is to reconcile contradictory specialist claims; an option designed to represent incomplete specialist work honestly would not directly provide it.
Question 342
A RAG search system is returning deleted documents and leaking restricted excerpts to an unauthorized user. Which TWO index controls address those failures? Choose TWO.
- Propagate source deletions and tombstones into the search index
- Carry source ACL metadata into the index and enforce retrieval filters
- Track search query latency, throttling and indexing resource saturation
- Retune OCR character confidence while continuing to serve deleted and unauthorized policy excerpts.
- Verify the index vector field dimensions match the embedding output
Explanation
Correct Answers
A:Documents removed from the source must be removed or deactivated in retrieval. Deletion propagation prevents stale evidence and ACL filtering protects private data at retrieval time.
B:Data permissions must be represented and applied before search results reach a caller.
Incorrect Answers
C:Index/search capacity and query complexity can cause tail latency independent of model generation. It would help when attempting to diagnose search infrastructure performance; the current failure mode is different: remove obsolete records and enforce source security.
D:OCR sampling cannot enforce source deletion or tenant-specific access. Reconcile deletions and apply trusted document permissions before any retrieval result is exposed.
E:Vector storage requires dimensional consistency between embeddings and the indexed vector field. Here, however, the stated objective is to remove obsolete records and enforce source security; this option focuses on resolve embedding dimensionality mismatch.
Question 343
A security analyst needs to locate an event at a known moment in a long recording. What preprocessing makes retrieval efficient?
- Index just the visual subject label for the entire recording and discard original time boundaries.
- Compare temporally sampled frames while preserving their order
- Segment the video into time-stamped intervals with searchable metadata
- Associate extracted frame features with source video IDs and segment offsets
Explanation
Correct Answer
C:Temporal chunking lets analysis and search refer to specific portions rather than treating video as one blob. This resolves the specific need to find an event in a large video by timestamp, while a measure that helps locate a detected subject in a frame addresses a different requirement.
Incorrect Answers
A:An event label without segment offsets cannot locate a known moment. Time-stamped video segments and searchable metadata preserve navigation back to the relevant footage.
B:Visual reasoning about motion requires correct temporal sequence, not unordered snapshots. That is relevant to compare observed visual changes through time, but the present scenario requires an ability to find an event in a large video by timestamp.
D:Provenance links allow retrieved frame evidence to be located in the original recording. The function serves to map search matches back to source video, whereas the actual requirement is to find an event in a large video by timestamp.
Question 344
A speech model repeatedly misunderstands specialized manufacturing acronyms. What customization is appropriate?
- Increase speech output volume without adjusting recognition configuration.
- Use supported custom speech vocabulary or phrase hints with representative evaluation audio.
- Translate the transcript into another language before correcting source recognition.
- Replace all unknown acronyms with the most common dictionary word.
Explanation
Correct Answer
B:Domain-specific acoustic and lexical terms may need adaptation and testing beyond a generic speech model.
Incorrect Answers
A:Audio playback loudness does not help the recognizer disambiguate technical acronyms.
C:Translation may compound a recognition error instead of fixing the original term.
D:Generic substitutions erase domain-specific terms and can corrupt manufacturing instructions.
Question 345
A video transcript and its image frames disagree about the displayed serial number. What evidence should be retained?
- Keep only the transcript and discard the corresponding source video frames.
- Preserve synchronized frame timestamps and transcript spans for reconciliation.
- Keep only the video frame without the time-aligned transcript evidence.
- Choose whichever modality has higher average benchmark accuracy globally.
Explanation
Correct Answer
B:Cross-modal evidence needs alignment; otherwise the extracted serial number cannot be audited to its source moment.
Incorrect Answers
A:Speech text alone cannot verify a serial number visibly printed in a particular frame.
C:The investigation requires comparing the spoken value and the displayed number at the same time.
D:Aggregate benchmarks do not resolve the specific conflicting serial evidence in this clip.
Question 346
A code-generation assistant passes syntax checks but fails to handle API timeouts. Which test should be added?
- Treat successful compilation as proof the generated client handles all network failures.
- Test only the happy path with a permanently responsive mock API.
- Increase logging detail without running failure-injection scenarios.
- Execute behavior-based tests with injected timeout and retry failures.
Explanation
Correct Answer
D:Parsing success does not prove robustness; fault-injection testing validates handling of operational errors.
Incorrect Answers
A:Syntax validation does not exercise timeout handling, retry behavior, or state recovery.
B:A normal response gives no evidence of correct behavior under dropped connections.
C:More logs may aid diagnosis but do not verify that the code behaves correctly on timeout.
Question 347
A containerized AI API stores a long-lived Foundry access key in environment variables. Which production authentication approach removes static secrets?
- Authenticate the workload with managed identity and supported Azure RBAC
- Keep the Foundry API key in the container environment and simply grant the same key a narrower resource role.
- Use authenticated developer identity with a supported credential chain and Azure role assignments
- Assign a narrowly scoped Storage Blob Data Reader role to its workload identity
Explanation
Correct Answer
A:Managed identity issues scoped tokens without storing resource keys in the app environment. The decisive difference is whether the design can replace embedded credentials with workload identity; the competing function serves to enforce least privilege for model inference.
Incorrect Answers
B:A leaked long-lived secret remains exploitable regardless of storage in an environment variable. A managed workload identity avoids distributing the reusable key and supports scoped RBAC.
C:Developer CLI-based identities can support local token authentication without distributing service keys. Its primary purpose is to use keyless local development credentials; the stated constraint instead demands a way to replace embedded credentials with workload identity.
D:Model invocation permission is distinct from reading blobs through the storage data plane. That is relevant to permit read-only blob data retrieval, but the present scenario requires an ability to replace embedded credentials with workload identity.
Question 348
An operations agent must describe what changes between the first and last frame of a clip. What workflow is needed?
- Locate the main object in only the first video frame and present its bounding box as the full temporal change summary.
- Store the video asset ID and frame timestamps but never compare the actual visual contents of the first and last frames.
- Sample a single middle frame, discard sequence order and infer both the initial and final states from that image.
- Compare temporally sampled frames while preserving their order
Explanation
Correct Answer
D:Visual reasoning about motion requires correct temporal sequence, not unordered snapshots. The practical requirement is to compare observed visual changes through time; a solution built to locate a detected subject in a frame would not meet that same need.
Incorrect Answers
A:A spatial bounding box answers where a subject appears at one instant, not what changed between two times. Preserve frame order and compare visual state across sampled frames to describe change.
B:Provenance allows locating a video frame but does not itself analyze movement or state changes. A temporal comparison must inspect both frames in their correct chronological order.
C:A middle snapshot cannot show the requested beginning-to-end change. Sample the relevant time positions and preserve temporal ordering, adjusting density when brief transitions matter.
Question 349
An organization tests a new custom speech model only on the audio used to train it. What evaluation fix is needed?
- Score a custom speech model only against the exact recordings it trained on
- Include representative background noise and device conditions in evaluation and adaptation
- Verify speaker consent, rights and permitted synthetic-voice usage
- Measure word error rate on a separate representative held-out audio set
Explanation
Correct Answer
D:Training data performance does not prove generalization to new speakers or noise. This resolves the specific need to evaluate speech recognition on unseen audio, while a measure that helps adapt ASR to industrial terminology addresses a different requirement.
Incorrect Answers
A:Training-set accuracy does not test generalization. Use held-out speakers and environments to assess performance before deployment.
B:Acoustic conditions strongly affect recognition, so realistic field recordings matter. That is relevant to validate recognition under real environmental noise, but the present scenario requires an ability to evaluate speech recognition on unseen audio.
C:Voice customization should not impersonate or clone real speakers without authorized rights. This helps when a team needs to obtain consent before creating a custom voice; it does not establish how to evaluate speech recognition on unseen audio.
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.