Microsoft AI-103 Practice Test 34
Question 330
Quality reviewers want to know whether an updated index improves search relevance. How should they compare versions?
- Judge retrieval quality only by the count of documents ingested
- Track extraction/OCR and enrichment errors separately from indexing success
- Run the same labeled retrieval benchmark against old and candidate index configurations
- Track indexing load and query latency and throttling together
Explanation
Correct Answer
C:Controlled comparisons distinguish relevance changes from differing query distributions. The practical requirement is to measure retrieval regressions across index changes; a solution built to detect ingestion completeness problems would not meet that same need.
Incorrect Answers
A:Index coverage does not establish that search results answer real queries. Evaluate relevance and answer grounding on labeled test questions.
B:Stage-specific health shows whether the problem lies in parsing, enrichment or index writes. This helps when a team needs to isolate which ingestion stage fails; it does not establish how to measure retrieval regressions across index changes.
D:Ingestion workload can compete with query capacity and affect interactive performance. That capability is appropriate for detect index refresh effects on search service, not this separate task of finding how to measure retrieval regressions across index changes.
Question 331
An incoming request needs only a simple exchange-rate lookup, yet every agent always launches. What routing logic is better?
- Run all specialist agents and aggregate every response for a straightforward exchange-rate lookup
- Launch the legal, coding, privacy and pricing specialists for every request, including a single exchange-rate lookup.
- Use a coordinating agent to assign specialists and consolidate sourced results
- Dispatch only the capabilities needed by the request
Explanation
Correct Answer
D:Selective orchestration avoids unnecessary model calls, latency and possible conflicting outputs. This resolves the specific need to avoid unnecessary agent fan-out, while a measure that helps represent incomplete specialist work honestly addresses a different requirement.
Incorrect Answers
A:Unnecessary agent fan-out adds cost and reconciliation work. A single authorized rate lookup is sufficient for the narrow request.
B:Unnecessary agents increase cost and failure paths. Dispatch only the capability needed for a simple lookup and escalate for genuine dependencies.
C:A central coordinator preserves coverage and resolves conflicts before one answer is returned. This helps when a team needs to coordinate specialists and synthesize results; it does not establish how to avoid unnecessary agent fan-out.
Question 332
Some private documents appear in anonymous search previews. What must the index pipeline verify?
- Sample OCR readability while leaving security labels out of the index and anonymous retrieval unrestricted.
- 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
Explanation
Correct Answer
C:Data permissions must be represented and applied before search results reach a caller. The practical requirement is to preserve document authorization during indexing; a solution built to check quality of OCR-derived searchable text would not meet that same need.
Incorrect Answers
A:Readable extracted text can still leak if document-level permissions are omitted. Carry trusted ACL metadata and enforce filtering for each caller before displaying excerpts.
B:Documents removed from the source must be removed or deactivated in retrieval. This helps when a team needs to remove obsolete indexed evidence; it does not establish how to preserve document authorization during indexing.
D:Index/search capacity and query complexity can cause tail latency independent of model generation. That capability is appropriate for diagnose search infrastructure performance, not this separate task of finding how to preserve document authorization during indexing.
Question 333
An intake solution analyzes both single-field photos and complex multi-stage documents, and must flag uncertain extractions. Which THREE design practices fit? Choose THREE.
- Accept extracted fields without retaining document layout, field-confidence evidence or review thresholds.
- Render all detected document regions at their raw coordinates on arbitrarily resized images without reconciling scale or orientation.
- Use a single-task Content Understanding analyzer configured for the required fields
- Evaluate pro-mode Content Understanding for a multi-stage analysis flow
- Flatten multi-column forms and tables into unordered plain text before extracting fields.
- Route low-confidence extracted fields to review before irreversible decisions
Explanation
Correct Answers
C:A focused extraction task can use a single-task pipeline without broad open-ended analysis. Single-task analysis, pro-mode for complex flows and low-confidence review match distinct workflow complexity and risk.
D:Advanced workflows may need a more capable orchestration mode than one fixed extraction.
F:Confidence and validation checks reduce high-impact errors from uncertain extraction.
Incorrect Answers
A:A multi-document workflow requires appropriate analyzer modes and uncertainty handling. Discarding structural evidence and low-confidence flags weakens all three controls.
B:Using incompatible coordinates can associate fields with the wrong region. This does not replace single-task or pro-mode extraction and independent low-confidence review.
E:Discarding layout relationships before extraction can misassociate amounts, labels and table rows.
Question 334
A factory recognition model routinely mistranscribes its private equipment names. What speech capability may help?
- Add more generic cloud microphone test samples but no recordings containing the factory's proprietary equipment vocabulary.
- Evaluate custom speech recognition with domain-specific labeled audio
- Measure word error rate on a separate representative held-out audio set
- Verify speaker consent, rights and permitted synthetic-voice usage
Explanation
Correct Answer
B:Custom speech can improve recognition for specialized vocabulary when adequate training data exists. The decisive difference is whether the design can adapt ASR to industrial terminology; the competing function serves to validate recognition under real environmental noise.
Incorrect Answers
A:Generic acoustic tests may miss recognition errors on specialized terms. Domain adaptation needs representative labeled utterances of actual equipment names and operating noise.
C:Training data performance does not prove generalization to new speakers or noise. Its primary purpose is to evaluate speech recognition on unseen audio; the stated constraint instead demands a way to adapt ASR to industrial terminology.
D:Voice customization should not impersonate or clone real speakers without authorized rights. That is relevant to obtain consent before creating a custom voice, but the present scenario requires an ability to adapt ASR to industrial terminology.
Question 335
An agent can summarize documents supplied in its prompt but cannot search the knowledge index. What integration is missing?
- Expose an authorized search tool with a bounded query interface to the agent
- Add a citation-formatting instruction to the agent prompt without exposing an authenticated search function in its available tools.
- Set result limits and relevance thresholds appropriate to the context budget
- Apply permission and tenant filters before returning search results to the model
Explanation
Correct Answer
A:Tool integration lets the agent retrieve relevant knowledge when needed without embedding entire corpora. Here the relevant capability must let the team enable an agent to query indexed evidence. An alternative targeting supply verifiable citation handles to an agent solves another problem.
Incorrect Answers
B:A prompt instruction cannot query a search index. The agent must receive an authorized retrieval tool with a bounded interface and relevant indexing credentials.
C:Bounded search responses preserve useful evidence while controlling tokens and noise. That capability is appropriate for limit and rank agent search-tool results, not this separate task of finding how to enable an agent to query indexed evidence.
D:RAG tools must honor caller identity and document ACLs rather than relying on the model to self-filter. Here the needed outcome is to enable an agent to query indexed evidence; an option designed to enforce search authorization for agent tools would not directly provide it.
Question 336
One of four delegated tasks fails but the coordinator reports full success. What lifecycle behavior is required?
- Track each task as pending, complete or failed and surface partial results accurately
- Mark the workflow successful once all prerequisite edges have been scheduled, even if one delegated agent failed.
- Mark every delegated task successful when any one specialist returns, even if another task failed.
- Give one owner write authority or serialize merges through the coordinator
Explanation
Correct Answer
A:Coordinators should not conceal a failed workstream or fabricate missing evidence. The practical requirement is to represent incomplete specialist work honestly; a solution built to run dependent steps in order and unrelated work in parallel would not meet that same need.
Incorrect Answers
B:Scheduling tasks is not the same as confirming successful completion. The coordinator must observe each specialist's status, handle errors or retries, and report any partial result explicitly.
C:Task completion must be tracked individually. Report failed and incomplete work explicitly rather than claiming an unsatisfied workflow is complete.
D:Concurrent stateful writes need ownership and a merge discipline to prevent lost changes. That capability is appropriate for avoid conflicting simultaneous agent edits, not this separate task of finding how to represent incomplete specialist work honestly.
Question 337
A query returns results but response latency rises with growing corpus size. What operational signal should be examined?
- Verify the index vector field dimensions match the embedding output
- Track search query latency, throttling and indexing resource saturation
- Measure precision at k on representative labeled questions
- Improve OCR character confidence and assume this will resolve the slow query plan as index size grows.
Explanation
Correct Answer
B:Index/search capacity and query complexity can cause tail latency independent of model generation. Here the relevant capability must let the team diagnose search infrastructure performance. An alternative targeting resolve embedding dimensionality mismatch solves another problem.
Incorrect Answers
A:Vector storage requires dimensional consistency between embeddings and the indexed vector field. This helps when a team needs to resolve embedding dimensionality mismatch; it does not establish how to diagnose search infrastructure performance.
C:Ranking precision tests whether the top results meet actual information needs. That capability is appropriate for evaluate search result relevance at the top, not this separate task of finding how to diagnose search infrastructure performance.
D:Text quality does not diagnose latency from indexing, query fan-out or ranking costs. Profile search latency, index load and resource capacity against representative traffic.
Question 338
A generated image contains a plausible but unlicensed company logo. What safeguard should be applied before publishing?
- Publish the image because the logo is slightly altered by generation.
- Apply brand and intellectual-property review with removal or licensed replacement of the mark.
- Remove the image's metadata and assume that clears branding rights.
- Rely exclusively on image-quality scoring to approve the mark.
Explanation
Correct Answer
B:Visual realism is not proof of permission to depict protected branding in a commercial asset.
Incorrect Answers
A:A recognizable unlicensed mark may still create trademark and usage-rights concerns.
C:Metadata removal has no bearing on the underlying visual trademark.
D:Aesthetics cannot establish whether an external company's logo is licensed for use.
Question 339
A brand wants a distinctive synthetic voice for customer service. What governance comes before voice creation?
- Verify speaker consent, rights and permitted synthetic-voice usage
- Clone a real employee's voice without explicit permissions or provenance controls
- Include representative background noise and device conditions in evaluation and adaptation
- Measure word error rate on a separate representative held-out audio set
Explanation
Correct Answer
A:Voice customization should not impersonate or clone real speakers without authorized rights. In this scenario the team must obtain consent before creating a custom voice. A feature used to adapt ASR to industrial terminology does not substitute for it.
Incorrect Answers
B:Synthetic voice creation requires authorization and misuse protections. Copying an individual's voice without consent bypasses governance safeguards.
C: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 obtain consent before creating a custom voice.
D:Training data performance does not prove generalization to new speakers or noise. This helps when a team needs to evaluate speech recognition on unseen audio; it does not establish how to obtain consent before creating a custom voice.
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.