Microsoft AI-103 Practice Test 29
Question 280
A generative model uses a high temperature for a regulated classification field and outputs unstable labels. What inference adjustment fits?
- Use lower-variance sampling and constrain labels through a supported schema or enum.
- Increase temperature further to encourage diverse category spellings.
- Remove the allowed-label enum and allow unrestricted explanatory prose.
- Change only the embedding index dimension, leaving generation settings unchanged.
Explanation
Correct Answer
A:Regulated classification benefits from repeatable outputs and explicit allowed labels instead of unbounded free-form generation.
Incorrect Answers
B:More randomness makes a regulated closed-label field less stable and harder to validate.
C:Removing the closed list of classification labels allows arbitrary text that downstream systems cannot validate. Keep an enum or equivalent schema constraint and test deterministic decoding settings on representative regulated examples.
D:Embedding dimensions control retrieval compatibility rather than classification output variance.
Question 281
The model's latency is stable but the share of unsafe responses rises for a new audience. Which telemetry should be tracked?
- Evaluate grounding faithfulness against newly curated reference examples
- Record only mean model latency and omit unsafe-response category and policy-violation counts.
- Track safety-filter outcomes and policy-violation rates by use case
- Track only incoming document format and extraction errors while ignoring safety-filter labels and verified unsafe completions.
Explanation
Correct Answer
C:Safety measurements reveal harmful output patterns that general request success metrics miss. The practical requirement is to observe content-policy violations; a solution built to monitor grounded factual correctness would not meet that same need.
Incorrect Answers
A:Quality checks must detect whether answers remain supported by changing authoritative evidence. That is relevant to monitor grounded factual correctness, but the present scenario requires an ability to observe content-policy violations.
B:A latency trend cannot reveal harmful content regression. Track use-case-specific safety outcomes and review confirmed policy violations.
D:Input-format drift may explain some failures, but the observed regression is an increase in unsafe model outputs. Measure policy violation categories, content-filter decisions and confirmed unsafe-response rates by audience.
Question 282
A write-capable agent must prevent malformed payment amounts, unauthorized high-value refunds and duplicate retries. Which THREE tool-level controls address these cases? Choose THREE.
- Validate submitted arguments against the function schema and reject the invalid call
- Replace all payment schema checks with generic access-denied errors and rely on model self-restraint for high-value refunds.
- Correlate each returned result to the original unique call identifier
- Check approval and value limits inside the application tool boundary before execution
- Expose all available write tools to the model, then assume that valid JSON alone guarantees permission for high-value refunds.
- Design side-effecting operations with idempotency and outcome reconciliation
Explanation
Correct Answers
A:Schema validation prevents nonsensical payloads reaching financial side effects. Schema validation, business authorization and idempotency cover distinct failure channels in tool execution.
D:Policy cannot depend on model compliance with prompt instructions when real money moves.
F:Safe retry requires recognizing requests already executed and not duplicating effects.
Incorrect Answers
B:Schema validation, trusted approval checks and idempotency prevent different risks. A blanket auth error conceals malformed input and does not make duplicate external writes safe.
C:The model execution protocol relies on stable identity for pending and completed tool operations. This addresses match tool results to the right call IDs. The scenario instead requires a way to validate inputs, enforce approval, prevent duplicate side effects.
E:Schema validity does not prove authority to issue a refund. The application must enforce limits and approval while making retries idempotent.
Question 283
A deployment update appears to change answer tone unexpectedly. How can operators attribute the regression?
- Correlate evaluation metrics with model, prompt and deployment version tags
- Track safety-filter outcomes and policy-violation rates by use case
- Compare input-format distribution and extraction-failure rates over time
- Discard prompt and deployment version labels before comparing the old and new answer-tone evaluations.
Explanation
Correct Answer
A:Version-labeled metrics distinguish configuration-induced drift from external user changes. Here the relevant capability must let the team attribute response drift to released versions. An alternative targeting observe content-policy violations solves another problem.
Incorrect Answers
B:Safety measurements reveal harmful output patterns that general request success metrics miss. The function serves to observe content-policy violations, whereas the actual requirement is to attribute response drift to released versions.
C:Shifting input characteristics can harm downstream model accuracy even without a new model release. That capability is appropriate for detect upstream data drift affecting model quality, not this separate task of finding how to attribute response drift to released versions.
D:Without release identifiers, a tone shift cannot be assigned to the changed model, prompt or application configuration. Retain versioned traces.
Question 284
A purchase-review agent receives typed Content Understanding fields with confidence scores and source-page spans. It must use those values for approval checks without confusing extracted table cells. What tool-result contract is most appropriate?
- Integrate visual extraction plus a separate authorized inventory lookup tool
- Expose a validated structured tool result retaining field types, confidence, table associations and source-page spans.
- Convert extracted tables into an untyped free-text paragraph and drop confidence and source-page references before approval.
- Flatten every extracted table into an unordered narrative and discard confidence scores.
Explanation
Correct Answer
B:A scanned-document extraction result should preserve typed fields, confidence and the page or region supporting every value. An agent or deterministic rule can then check totals, handle low-confidence fields and trace decisions to the original document. Unordered prose is a weaker contract for financial approvals. Typed and source-linked extraction lets downstream rules validate amounts and confidence before acting, rather than trusting a free-form narrative about a scanned form.
Incorrect Answers
A:The image supplies observable item details; the live database supplies current availability. This is a separate capability and does not directly consume confidence- and provenance-aware structured extraction as typed agent tool data.
C:Typed fields, table associations and evidence spans are necessary for reliable checks. Flattening loses the identity and reliability of the extracted values.
D:Dropping field types, row associations and confidence values makes numerical approval checks less reliable, especially when extraction is uncertain.
Question 285
Users complain that answers omit specific evidence, though retrieval returns relevant passages. What application-level metric matters?
- Track only median model latency and treat lower response time as proof that all requested evidence was cited.
- Classify errors by validation, authorization, rate-limit and service fault categories
- Compare input-format distribution and extraction-failure rates over time
- Measure citation coverage and grounded-answer relevance on a labeled test set
Explanation
Correct Answer
D:Grounding-quality evaluation detects failures between successful retrieval and supported generation. This directly supports the need to monitor evidence use in generated answers, unlike a mechanism primarily intended to attribute response drift to released versions.
Incorrect Answers
A:Latency measures performance, not faithfulness or citation coverage. Evaluate whether grounded answers contain required evidence and accurately cite the retrieved passages.
B:Structured error codes enable targeted remediation instead of assuming every failed request requires retries. That capability is appropriate for separate service errors by actionable failure type, not this separate task of finding how to monitor evidence use in generated answers.
C:Shifting input characteristics can harm downstream model accuracy even without a new model release. Here the needed outcome is to monitor evidence use in generated answers; an option designed to detect upstream data drift affecting model quality would not directly provide it.
Question 286
A team modifies both temperature and top-p during a quality experiment and cannot attribute the observed change. What evaluation design helps?
- Change temperature and top-p together again, then compare one successful example.
- Change one sampling setting at a time against a fixed, versioned evaluation set.
- Use different validation questions in each experiment without recording the differences.
- Select the preferred setting solely from one subjective response without controlled scoring.
Explanation
Correct Answer
B:Controlled comparisons isolate parameter effects and reduce confounding from multiple simultaneous changes.
Incorrect Answers
A:Multiple simultaneous changes and a single sample cannot isolate the cause of the behavior shift.
C:Changing test inputs introduces an additional confound into the sampling comparison.
D:Choosing sampling parameters from one unscored answer confounds model variation with actual quality improvement. Hold the evaluation cases and other settings constant, change one parameter at a time and score repeated outputs.
Question 287
A chatbot returns more apologies after customer-upload formats change. What distribution signal might expose the source?
- Monitor only GPU utilization and ignore the newly introduced file formats and extraction-failure patterns.
- Correlate evaluation metrics with model, prompt and deployment version tags
- Evaluate grounding faithfulness against newly curated reference examples
- Compare input-format distribution and extraction-failure rates over time
Explanation
Correct Answer
D:Shifting input characteristics can harm downstream model accuracy even without a new model release. The decisive difference is whether the design can detect upstream data drift affecting model quality; the competing function serves to observe content-policy violations.
Incorrect Answers
A:Compute utilization cannot identify a shift toward unsupported or poorly parsed upload types. Measure format mix, parsing failures and input-quality drift alongside downstream response outcomes.
B:Version-labeled metrics distinguish configuration-induced drift from external user changes. Here the needed outcome is to detect upstream data drift affecting model quality; an option designed to attribute response drift to released versions would not directly provide it.
C:Quality checks must detect whether answers remain supported by changing authoritative evidence. Its primary purpose is to monitor grounded factual correctness; the stated constraint instead demands a way to detect upstream data drift affecting model quality.
Question 288
A claims agent receives scanned forms whose fields are not searchable as text. Which extraction tool belongs in its workflow?
- Attach an authorized retrieval tool that provides source passages and identifiers
- Query only indexed image filenames and skip text-and-layout extraction from the scanned claims forms.
- Invoke layout-aware Content Understanding to extract fields from scanned forms
- Integrate visual extraction plus a separate authorized inventory lookup tool
Explanation
Correct Answer
C:Document analyzers recover text, table structure and fields before downstream reasoning. The practical requirement is to extract structured content from scanned claim forms; a solution built to ground agent answers in accessible policy passages would not meet that same need.
Incorrect Answers
A:A retrieval tool lets the agent ground responses in current corporate documents and cite the actual evidence. That is relevant to ground agent answers in accessible policy passages, but the present scenario requires an ability to extract structured content from scanned claim forms.
B:Names of image files cannot provide fields printed in their pixels. Invoke an appropriate document extraction tool before searching or validating the fields.
D:The image supplies observable item details; the live database supplies current availability. That capability is appropriate for combine visual item facts and live stock data, not this separate task of finding how to extract structured content from scanned claim forms.
Question 289
A retrieval service suddenly returns irrelevant hits after its vector index was rebuilt. What should the operator inspect?
- Compare embedding model, vector dimensionality, index schema, and query embedding configuration.
- Increase the search top-k indefinitely without checking the vectors.
- Rebuild only the index's display name while leaving incompatible query embeddings.
- Disable retrieval logs so the relevance regression no longer appears.
Explanation
Correct Answer
A:Vectors from different embedding spaces may share a numeric length but not semantic compatibility. Rebuild and query with a compatible representation.
Incorrect Answers
B:Retrieving more results cannot make incompatible query and document embeddings comparable.
C:Renaming an index does not repair vector dimensional or model-space incompatibility.
D:Hiding telemetry does not correct the index schema or embedding configuration.
Question 290
An agent needs a live shipping estimate from an authenticated carrier API. How should that fact be obtained?
- Invoke layout-aware Content Understanding to extract fields from scanned forms
- Estimate the shipment date from old conversational examples instead of calling the authorized carrier endpoint.
- Use a narrowly scoped shipping API tool instead of reasoning from stored training text
- Parse typed Content Understanding output and preserve table structure as data
Explanation
Correct Answer
C:Delivery estimates change; authorized live tool data is more trustworthy than historical model memory. Here the relevant capability must let the team retrieve current carrier estimates through an API. An alternative targeting extract structured content from scanned claim forms solves another problem.
Incorrect Answers
A:Document analyzers recover text, table structure and fields before downstream reasoning. This helps when a team needs to extract structured content from scanned claim forms; it does not establish how to retrieve current carrier estimates through an API.
B:A live estimate depends on current carrier data. A model should invoke an authenticated, suitably scoped shipping API instead of inventing an ETA.
D:Structured parsing distinguishes extracted values from instructions and maintains cell relationships. Here the needed outcome is to retrieve current carrier estimates through an API; an option designed to consume structured document extraction output reliably would not directly provide it.