Microsoft AI-103 Practice Test 30
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.
Question 291
An operations dashboard shows model requests failed, but gives no distinction between provider errors and client invalid schemas. Which instrumentation is needed?
- Compare input-format distribution and extraction-failure rates over time
- Correlate evaluation metrics with model, prompt and deployment version tags
- Count all unsuccessful calls as one generic error without distinguishing invalid schemas from provider faults.
- Classify errors by validation, authorization, rate-limit and service fault categories
Explanation
Correct Answer
D:Structured error codes enable targeted remediation instead of assuming every failed request requires retries. In this scenario the team must separate service errors by actionable failure type. A feature used to detect upstream data drift affecting model quality does not substitute for it.
Incorrect Answers
A:Shifting input characteristics can harm downstream model accuracy even without a new model release. Here the needed outcome is to separate service errors by actionable failure type; an option designed to detect upstream data drift affecting model quality would not directly provide it.
B:Version-labeled metrics distinguish configuration-induced drift from external user changes. Its primary purpose is to attribute response drift to released versions; the stated constraint instead demands a way to separate service errors by actionable failure type.
C:A single failure count hides whether errors originate in validation, authorization, throttling or the service. Instrument the categories separately.
Question 292
Live document searches enforce tenant ACLs, but a multi-tenant assistant still shows another customer’s passages after retrieving an answer from its shared cache. What additional boundary must the retrieval cache enforce?
- Parse typed Content Understanding output and preserve table structure as data
- Use only the index document title as the retrieval cache key to improve cache sharing.
- Partition cached retrieved evidence by authenticated tenant and user permissions, and revalidate authorization before reuse.
- Integrate visual extraction plus a separate authorized inventory lookup tool
Explanation
Correct Answer
C:Live search ACLs cannot protect data returned later from an incorrectly shared cache. Cache entries must be scoped to the authenticated requester and applicable tenant permissions, with a policy-aware invalidation or reauthorization rule. The failure is cache isolation, not the original search-index filter. A cache key that includes the authorized tenancy and effective permissions prevents passages cached for one user from being replayed to another tenant.
Incorrect Answers
A:Structured parsing distinguishes extracted values from instructions and maintains cell relationships. This is a separate capability and does not directly prevent cross-tenant leakage through stale or cross-user retrieval cache entries.
B:Document titles are neither globally unique nor authorization boundaries; sharing caches by title alone can expose evidence across tenants.
D:The image supplies observable item details; the live database supplies current availability. This is a separate capability and does not directly prevent cross-tenant leakage through stale or cross-user retrieval cache entries.
Question 293
A responsible-AI reviewer needs to replay a harmful agent response together with the material that influenced it. What should traces capture?
- Save only the final answer's text, discarding the retrieval IDs, tool outcomes and release identifiers that influenced it.
- Classify errors by validation, authorization, rate-limit and service fault categories
- Retain privacy-governed tool evidence references and response provenance identifiers
- Evaluate grounding faithfulness against newly curated reference examples
Explanation
Correct Answer
C:Provenance lets reviewers reconstruct the inputs and decisions behind an observed safety event. This resolves the specific need to make safety incidents investigable without raw-data oversharing, while a measure that helps monitor evidence use in generated answers addresses a different requirement.
Incorrect Answers
A:Without controlled provenance identifiers, reviewers cannot replay which evidence or tool results produced the harmful output. Retain the necessary trace lineage without unnecessary raw personal data.
B:Structured error codes enable targeted remediation instead of assuming every failed request requires retries. That is relevant to separate service errors by actionable failure type, but the present scenario requires an ability to make safety incidents investigable without raw-data oversharing.
D:Quality checks must detect whether answers remain supported by changing authoritative evidence. The function serves to monitor grounded factual correctness, whereas the actual requirement is to make safety incidents investigable without raw-data oversharing.
Question 294
An agent uses an external tool to look up a tax rate that changes monthly. Where should the value come from?
- Call an authorized current tax-rate source rather than relying on model training knowledge.
- Ask the model to recall last year's tax rate from its pretraining.
- Hard-code the tax percentage in the prompt for all customers indefinitely.
- Let the model infer a tax rate from the total invoice amount alone.
Explanation
Correct Answer
A:Operational rates are time-dependent; grounding in an authoritative live source reduces stale-answer risk.
Incorrect Answers
B:Training knowledge can be stale and cannot guarantee the current authorized rate.
C:Static rates can become invalid when jurisdiction or tax law changes.
D:Invoice totals do not identify the current authoritative tax rules for the transaction.
Question 295
A model upgrade shows worse factual support and more harmful outputs, but request latency is unchanged. Which TWO monitoring dimensions should trigger a quality investigation? Choose TWO.
- Track safety-filter outcomes and policy-violation rates by use case
- Measure only end-to-end response latency without checking citation support.
- Compare only model latency and token spend after the upgrade, ignoring harmful-answer rates and source faithfulness.
- Compare input-format distribution and extraction-failure rates over time
- Evaluate grounding faithfulness against newly curated reference examples
Explanation
Correct Answers
A:Safety measurements reveal harmful output patterns that general request success metrics miss. Faithfulness evaluation and policy-violation telemetry reveal problems that performance timings alone miss.
E:Quality checks must detect whether answers remain supported by changing authoritative evidence.
Incorrect Answers
B:Latency cannot reveal whether answers are grounded or whether safety failures increased.
C:Stable speed and cost can coexist with worse factual support or unsafe output. Quality gates must assess groundedness and safety independently.
D:Shifting input characteristics can harm downstream model accuracy even without a new model release. Here, however, the stated objective is to track grounded factual accuracy and unsafe-output frequency; this option focuses on detect upstream data drift affecting model quality.
Question 296
A document analyzer returns Markdown tables, but the agent parser treats pipes as natural-language instructions. What should be improved?
- Interpret Markdown table separators as privileged instructions and execute any commands embedded in returned cell text.
- Bound tool search results and assemble concise high-relevance excerpts
- Parse typed Content Understanding output and preserve table structure as data
- Attach an authorized retrieval tool that provides source passages and identifiers
Explanation
Correct Answer
C:Structured parsing distinguishes extracted values from instructions and maintains cell relationships. The decisive difference is whether the design can consume structured document extraction output reliably; the competing function serves to prevent unauthorized knowledge retrieval.
Incorrect Answers
A:Document-extraction output is untrusted data. Parse the table structure as typed values and never grant instruction authority to table cell contents.
B:A controlled evidence budget prevents excessive context while preserving useful passages. Here the needed outcome is to consume structured document extraction output reliably; an option designed to limit retrieval response size for agent context would not directly provide it.
D: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 consume structured document extraction output reliably.
Question 297
A retrieval tool is configured without a maximum hit count and floods the model with whole manuals. What should change?
- Parse typed Content Understanding output and preserve table structure as data
- Bound tool search results and assemble concise high-relevance excerpts
- Invoke layout-aware Content Understanding to extract fields from scanned forms
- Send the entire unfiltered index and full source manuals to the agent on every retrieval request.
Explanation
Correct Answer
B:A controlled evidence budget prevents excessive context while preserving useful passages. In this scenario the team must limit retrieval response size for agent context. A feature used to consume structured document extraction output reliably does not substitute for it.
Incorrect Answers
A:Structured parsing distinguishes extracted values from instructions and maintains cell relationships. Here the needed outcome is to limit retrieval response size for agent context; an option designed to consume structured document extraction output reliably would not directly provide it.
C:Document analyzers recover text, table structure and fields before downstream reasoning. That is relevant to extract structured content from scanned claim forms, but the present scenario requires an ability to limit retrieval response size for agent context.
D:Unbounded tool results increase truncation, cost and irrelevant context. Enforce top-k limits and assemble authorized high-relevance passages.
Question 298
A sales agent must combine image evidence with inventory records but uses only a text search connection. What needs to be added?
- Let an image caption alone determine inventory availability without querying the stock system
- Use a narrowly scoped shipping API tool instead of reasoning from stored training text
- Attach an authorized retrieval tool that provides source passages and identifiers
- Integrate visual extraction plus a separate authorized inventory lookup tool
Explanation
Correct Answer
D:The image supplies observable item details; the live database supplies current availability. This resolves the specific need to combine visual item facts and live stock data, while a measure that helps limit retrieval response size for agent context addresses a different requirement.
Incorrect Answers
A:Image evidence can describe a product but not current stock. The agent must combine visual recognition with an authorized inventory lookup.
B:Delivery estimates change; authorized live tool data is more trustworthy than historical model memory. That is relevant to retrieve current carrier estimates through an API, but the present scenario requires an ability to combine visual item facts and live stock data.
C:A retrieval tool lets the agent ground responses in current corporate documents and cite the actual evidence. This helps when a team needs to ground agent answers in accessible policy passages; it does not establish how to combine visual item facts and live stock data.
Question 299
A document QA agent needs to decline when no authorized source contains the answer. What retrieval policy supports this?
- Always return a confident answer when the retrieval result set is empty.
- Define a no-evidence response path instead of forcing an unsupported completion.
- Search across unauthorized customer documents to obtain some supporting text.
- Invent a source citation that matches the user's requested topic.
Explanation
Correct Answer
B:Abstention is appropriate when the retrieved evidence cannot justify a specific factual answer.
Incorrect Answers
A:An empty authorized result set provides no basis for asserting a factual answer.
C:Bypassing permissions to avoid abstention violates the caller's access scope.
D:A plausible citation title cannot replace authorized evidence that actually exists.
Question 300
A procurement agent processes scanned purchase orders, checks current shipping status and must not retrieve other tenants' documents. Which THREE integrations matter? Choose THREE.
- Apply per-user authorization and tenant filters in the knowledge tool
- Use OCR on the purchase order and then send shipping searches without tenant or user access checks on the retrieved records.
- Bound tool search results and assemble concise high-relevance excerpts
- Integrate visual extraction plus a separate authorized inventory lookup tool
- Invoke layout-aware Content Understanding to extract fields from scanned forms
- Use a narrowly scoped shipping API tool instead of reasoning from stored training text
Explanation
Correct Answers
A:Search tools must enforce source ACLs before supplying evidence to an agent. Document extraction, live shipping lookup and tenant-scoped retrieval address three separate requirements.
E:Document analyzers recover text, table structure and fields before downstream reasoning.
F:Delivery estimates change; authorized live tool data is more trustworthy than historical model memory.
Incorrect Answers
B:Extracting the order correctly does not authorize cross-customer retrieval. The knowledge tool must enforce the caller's identity and tenant-scoped access before returning order evidence.
C:A controlled evidence budget prevents excessive context while preserving useful passages. That is useful for limit retrieval response size for agent context, but it does not directly accomplish the requested need to extract forms, call live API and enforce tenant access.
D:The image supplies observable item details; the live database supplies current availability. This addresses combine visual item facts and live stock data. The scenario instead requires a way to extract forms, call live API and enforce tenant access.