Microsoft AI-103 Practice Test 33
Question 320
A Foundry Content Understanding analyzer emits clean Markdown, but the business app needs typed invoice fields. What should be configured?
- Use a plain Markdown output without any field definitions and parse it heuristically.
- Increase the OCR resolution while retaining untyped narrative-only output.
- Ask the agent to guess column names from whichever invoice arrives next.
- Select or define an analyzer output schema that includes the required structured fields.
Explanation
Correct Answer
D:Markdown captures readable content; a downstream accounting interface may require a validated field schema.
Incorrect Answers
A:Markdown layout may be readable but does not guarantee typed invoice fields for the application.
B:Better character recognition does not define a structured output schema for fields.
C:Uncontrolled naming produces unstable field contracts and breaks downstream processing.
Question 321
A coordinator gets a completed specialist report with no evidence identifiers. What result contract should be required?
- Require source references, scope and confidence in each specialist deliverable
- Give one owner write authority or serialize merges through the coordinator
- Use a coordinating agent to assign specialists and consolidate sourced results
- Order the specialists by prerequisite dependencies but allow reports to omit evidence identifiers and source references.
Explanation
Correct Answer
A:Traceable outputs let the coordinating agent verify and cite claims instead of merging unsupported summaries. In this scenario the team must make specialist results auditable. A feature used to avoid conflicting simultaneous agent edits does not substitute for it.
Incorrect Answers
B:Concurrent stateful writes need ownership and a merge discipline to prevent lost changes. Its primary purpose is to avoid conflicting simultaneous agent edits; the stated constraint instead demands a way to make specialist results auditable.
C:A central coordinator preserves coverage and resolves conflicts before one answer is returned. That is relevant to coordinate specialists and synthesize results, but the present scenario requires an ability to make specialist results auditable.
D:A dependency graph controls timing, not evidence traceability. Require structured outputs that identify the observations and source references so the coordinator can validate the specialist's claims.
Question 322
The top-ranked search result is often unrelated despite healthy indexing. Which quality check is most informative?
- Verify the index vector field dimensions match the embedding output
- Re-run OCR confidence checks and assume correct text extraction alone guarantees the highest-ranked result is relevant.
- Propagate source deletions and tombstones into the search index
- Measure precision at k on representative labeled questions
Explanation
Correct Answer
D:Ranking precision tests whether the top results meet actual information needs. In this scenario the team must evaluate search result relevance at the top. A feature used to resolve embedding dimensionality mismatch does not substitute for it.
Incorrect Answers
A:Vector storage requires dimensional consistency between embeddings and the indexed vector field. Here the needed outcome is to evaluate search result relevance at the top; an option designed to resolve embedding dimensionality mismatch would not directly provide it.
B:OCR quality is only an upstream ingestion check. Relevance requires query-result judgments, recall/ranking metrics and representative graded searches.
C:Documents removed from the source must be removed or deactivated in retrieval. That is relevant to remove obsolete indexed evidence, but the present scenario requires an ability to evaluate search result relevance at the top.
Question 323
A service wants to audit how often visual extraction misses safety labels on low-resolution photos. What evaluator is appropriate?
- Evaluate object-box coordinate precision but never count the safety labels that the model failed to detect.
- Retain layout and table-cell relationships rather than flattening all OCR words
- Route low-confidence extracted fields to review before irreversible decisions
- Measure label-detection recall using reviewed annotated images
Explanation
Correct Answer
D:Detection recall on representative inputs reveals missed objects important to the workflow. This directly supports the need to detect missed visual labels quantitatively, unlike a mechanism primarily intended to align extracted regions with rendered images.
Incorrect Answers
A:Precise boxes for detected objects do not measure missed labels. Recall against reviewed annotated images captures false negatives, especially for small low-resolution warnings.
B:Structured layout informs which values belong to which headings and rows. That capability is appropriate for preserve table structure from document images, not this separate task of finding how to detect missed visual labels quantitatively.
C:Confidence and validation checks reduce high-impact errors from uncertain extraction. Here the needed outcome is to detect missed visual labels quantitatively; an option designed to escalate uncertain extracted visual fields would not directly provide it.
Question 324
A multilingual service must switch voice style with the user's chosen language. How should the voice be selected?
- Choose a compatible language and voice configuration for synthesis
- Ignore the chosen output language and reuse the previously selected voice
- Use text-to-speech synthesis for the response text
- Apply an appropriate retention and caching policy to synthesized audio
Explanation
Correct Answer
A:An appropriate voice/language match improves pronunciation and user experience. Here the relevant capability must let the team synthesize speech in the selected language. An alternative targeting improve spoken output pacing solves another problem.
Incorrect Answers
B:Voice locale should match the user's language and supported voice profile. Carrying forward an incompatible voice can degrade pronunciation and accessibility.
C:Speech synthesis converts generated or predetermined text into audio suitable for playback. That capability is appropriate for turn agent text into spoken output, not this separate task of finding how to synthesize speech in the selected language.
D:Retention should reflect use case, privacy and reuse rather than automatically saving all speech indefinitely. Here the needed outcome is to synthesize speech in the selected language; an option designed to manage storage and privacy for synthesized replies would not directly provide it.
Question 325
A search service slows down during a document refresh and query timeouts rise. What should be monitored?
- Reindex every document synchronously during the busiest query window with no capacity control
- Track extraction/OCR and enrichment errors separately from indexing success
- Track indexing load and query latency and throttling together
- Run the same labeled retrieval benchmark against old and candidate index configurations
Explanation
Correct Answer
C:Ingestion workload can compete with query capacity and affect interactive performance. This resolves the specific need to detect index refresh effects on search service, while a measure that helps detect ingestion completeness problems addresses a different requirement.
Incorrect Answers
A:Unbounded concurrent refresh can compete with interactive query resources. Throttling, scheduling and capacity observation are needed.
B:Stage-specific health shows whether the problem lies in parsing, enrichment or index writes. That is relevant to isolate which ingestion stage fails, but the present scenario requires an ability to detect index refresh effects on search service.
D:Controlled comparisons distinguish relevance changes from differing query distributions. The function serves to measure retrieval regressions across index changes, whereas the actual requirement is to detect index refresh effects on search service.
Question 326
An agent evaluator reports a correct final answer even though it used a disallowed customer-data export tool. What else must be measured?
- Evaluate the complete tool trajectory and compliance with permission policies.
- Judge only the final text's factual correctness and ignore how it was obtained.
- Review just the duration of the run as a safety metric.
- Trust the model's self-report that no restricted tool was invoked.
Explanation
Correct Answer
A:Outcome accuracy cannot justify a prohibited action. Trace evaluation needs to inspect intermediate behavior.
Incorrect Answers
B:An agent can arrive at the right answer by executing prohibited tools or accessing restricted data.
C:Runtime does not reveal which permissions or tool paths were used.
D:A generated assurance is not a trace-based audit of actual tool calls.
Question 327
A vector index begins rejecting newly embedded documents after the embedding model changes. Which schema issue should be checked?
- Verify the index vector field dimensions match the embedding output
- Raise the OCR confidence threshold to fix a vector field rejecting embeddings with the wrong dimension count.
- Measure precision at k on representative labeled questions
- Delete the document ACL metadata field to resolve mismatched embedding dimensions in the vector index.
Explanation
Correct Answer
A:Vector storage requires dimensional consistency between embeddings and the indexed vector field. This resolves the specific need to resolve embedding dimensionality mismatch, while a measure that helps check quality of OCR-derived searchable text addresses a different requirement.
Incorrect Answers
B:OCR confidence does not control the embedding vector's declared length. Verify model output dimensions and rebuild or migrate index schema as needed.
C:Ranking precision tests whether the top results meet actual information needs. This helps when a team needs to evaluate search result relevance at the top; it does not establish how to resolve embedding dimensionality mismatch.
D:Access metadata and vector dimensions serve different purposes. Update the embedding-field schema/model compatibility without weakening document authorization.
Question 328
A document analyzer emits confidence signals for detected fields but the application ignores them. What downstream rule helps?
- Automatically approve every extracted financial field whenever the analyzer returns syntactically valid JSON, regardless of confidence.
- Configure a visual analyzer to identify relevant objects and regions
- Evaluate pro-mode Content Understanding for a multi-stage analysis flow
- Route low-confidence extracted fields to review before irreversible decisions
Explanation
Correct Answer
D:Confidence and validation checks reduce high-impact errors from uncertain extraction. The decisive difference is whether the design can escalate uncertain extracted visual fields; the competing function serves to align extracted regions with rendered images.
Incorrect Answers
A:A well-formed field can still be wrong. High-stakes downstream actions require confidence thresholds, cross-checks and review for uncertain extracted amounts or identities.
B:Visual extraction can return object-level evidence for downstream validation or action. Here the needed outcome is to escalate uncertain extracted visual fields; an option designed to extract object and region information from an image would not directly provide it.
C:Advanced workflows may need a more capable orchestration mode than one fixed extraction. Its primary purpose is to run a richer multimodal understanding pipeline; the stated constraint instead demands a way to escalate uncertain extracted visual fields.
Question 329
A voice app stores every generated reply as audio even when users only need one playback. What operational improvement helps?
- Persist each synthesized audio response indefinitely even after the associated session expires
- Choose a compatible language and voice configuration for synthesis
- Apply an appropriate retention and caching policy to synthesized audio
- Control the supported speech rate and pronunciation settings
Explanation
Correct Answer
C:Retention should reflect use case, privacy and reuse rather than automatically saving all speech indefinitely. This directly supports the need to manage storage and privacy for synthesized replies, unlike a mechanism primarily intended to turn agent text into spoken output.
Incorrect Answers
A:Unlimited storage increases privacy and retention risk without supporting users who only need text. Apply modality- and purpose-specific retention.
B:An appropriate voice/language match improves pronunciation and user experience. That capability is appropriate for synthesize speech in the selected language, not this separate task of finding how to manage storage and privacy for synthesized replies.
D:Speech rendering controls can improve intelligibility without altering the underlying facts. Its primary purpose is to improve spoken output pacing; the stated constraint instead demands a way to manage storage and privacy for synthesized replies.
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.