Microsoft AI-103 Practice Test 17
Question 160
A semantic search prototype uses only one paraphrased query as its benchmark. What evaluation improvement matters?
- Declare search quality satisfactory based on a single paraphrase and no representative negatives
- Benchmark the retriever on one hand-picked query and report success without testing short queries, jargon, synonyms or paraphrases.
- Measure precision at five against labeled relevant passages
- Build a diverse labeled query set covering short and long phrasing
Explanation
Correct Answer
D:Relevance assessment needs representative query diversity before claiming robust semantic quality. This directly supports the need to evaluate search across different query phrasings, unlike a mechanism primarily intended to retrieve paraphrased source passages.
Incorrect Answers
A:One successful query cannot measure robustness to wording variation or false matches. A held-out set of representative queries and hard negatives is needed.
B:A single favorable phrasing cannot establish retrieval robustness. Build a diverse labeled query set representing the real language variations users will submit.
C:Precision at k directly measures how often top results meet user information needs. Here the needed outcome is to evaluate search across different query phrasings; an option designed to score accuracy of top search results would not directly provide it.
Question 161
An agent is expected to call a shipping API, but the model invents an argument property not in the tool contract. Which safeguard fits?
- Forward the unrecognized argument property and let the shipping backend decide.
- Treat the model-generated JSON as authoritative whenever it parses.
- Auto-create missing backend fields based on the model's guessed property name.
- Validate tool-call arguments against the registered schema before executing the backend request.
Explanation
Correct Answer
D:The trusted tool executor must reject missing, extra or invalid fields rather than assuming model-generated JSON is correct.
Incorrect Answers
A:Unexpected tool parameters can trigger unintended operations or cause schema and security failures.
B:JSON syntax does not guarantee compliance with the registered tool schema.
C:Guessing a schema extension bypasses the reviewed contract and increases injection risk.
Question 162
Several apps need to consume one deployed model, but individual teams should not share privileged administrator credentials. What identity approach fits?
- Use separate appropriately governed projects and resource scopes for environments
- Distribute model controllers across zones while giving every app the same administrator credential.
- Load test end-to-end model calls, tool latency and downstream concurrency limits
- Give each calling workload its own managed identity and scoped role
Explanation
Correct Answer
D:Distinct workload identities permit least-privilege access without distributing a shared secret. This resolves the specific need to isolate each calling service's access rights, while a measure that helps isolate development and production resources addresses a different requirement.
Incorrect Answers
A:Environment separation limits accidental cross-environment changes and supports distinct access and cost controls. Its primary purpose is to isolate development and production resources; the stated constraint instead demands a way to isolate each calling service's access rights.
B:Controller availability does not establish per-application authorization. Each workload should authenticate with its own managed identity or service principal and the narrow role needed for its model invocation.
C:AI scalability depends on model throughput and tool dependencies, not only the web server instance count. This helps when a team needs to validate full pipeline capacity under a peak; it does not establish how to isolate each calling service's access rights.
Question 163
A warehouse agent needs to know whether a safety helmet is visible at a particular point in a video. What evidence format is most reliable?
- Return only a yes or no without time or source frame information.
- Cite the video filename but not the inspected segment.
- Return the observed frame timestamp and the visual region supporting the finding.
- Infer helmet visibility from an audio transcript without checking the frame.
Explanation
Correct Answer
C:Timestamped, localized visual evidence is more defensible than a general ungrounded claim about the entire clip.
Incorrect Answers
A:A bare label cannot show which portion of a video supports the helmet observation.
B:A filename is insufficient for verifying a brief event at a particular timestamp.
D:Speech content does not establish the visual presence of protective equipment.
Question 164
A support center must translate customer messages into Spanish with consistent technical terminology. Which capability should it use?
- Use Azure Translator with a managed terminology strategy where supported
- Detect the source language before choosing the translation direction
- Request faithful translation only and validate that no extra commentary was added
- Preserve and compare numeric, currency and identifier tokens across translations
Explanation
Correct Answer
A:A dedicated translation flow can provide multilingual conversion while preserving approved terminology. This resolves the specific need to translate support text with consistent terminology, while a measure that helps select the correct source language automatically addresses a different requirement.
Incorrect Answers
B:Language identification prevents using the wrong source and target settings. That is relevant to select the correct source language automatically, but the present scenario requires an ability to translate support text with consistent terminology.
C:Translation output should preserve meaning and scope rather than introduce advisory material. This helps when a team needs to avoid additions when translating legal text; it does not establish how to translate support text with consistent terminology.
D:Translation must not silently alter financial facts or reference numbers. The function serves to protect exact business values during translation, whereas the actual requirement is to translate support text with consistent terminology.
Question 165
A troubleshooting bot must retrieve an exact error code and conceptually related symptoms from the same index. What strategy fits?
- Search only semantic embeddings and drop exact error-code tokens from the query and index.
- Use compatible embedding representations for indexed documents and search queries
- Align embedding output dimension with the index vector field definition
- Use hybrid keyword and vector retrieval in one query workflow
Explanation
Correct Answer
D:Hybrid retrieval covers rare identifiers and semantic descriptions that neither method handles perfectly alone. The decisive difference is whether the design can match exact identifiers plus related concepts; the competing function serves to avoid redundant evidence in hybrid retrieval.
Incorrect Answers
A:Vector similarity helps conceptual matches, but exact error strings need lexical matching. A hybrid query preserves code-level precision while also retrieving conceptually related symptoms.
B:Vector distances are meaningful only when query and document vectors inhabit compatible spaces. Here the needed outcome is to match exact identifiers plus related concepts; an option designed to keep document and query embeddings compatible would not directly provide it.
C:Vector search requires vectors whose shape agrees with the configured schema. Its primary purpose is to prevent vector field shape mismatches; the stated constraint instead demands a way to match exact identifiers plus related concepts.
Question 166
A prompt revision improves evaluation scores on 20 handcrafted examples but fails on new user phrasing. What should be strengthened?
- Evaluate the agent tool trajectory and policy compliance, not just final text
- Retest only the twenty prompt examples used during tuning and treat their score increase as deployment proof.
- Use separate representative validation data and regression holdout cases
- Measure task fulfillment and relevance against representative user goals
Explanation
Correct Answer
C:Evaluating only on examples used to tune a prompt creates overfitting and misleading confidence. In this scenario the team must test generalization beyond tuning cases. A feature used to evaluate tool-call behavior for safety does not substitute for it.
Incorrect Answers
A:A successful final answer does not justify unauthorized intermediate actions. Here the needed outcome is to test generalization beyond tuning cases; an option designed to evaluate tool-call behavior for safety would not directly provide it.
B:Reusing the tuning set risks selection bias. A separate holdout with diverse phrasing and failure cases is needed to detect regressions and unsupported generalization.
D:Keyword coverage alone does not prove the answer resolves the user's problem. This helps when a team needs to evaluate real task success; it does not establish how to test generalization beyond tuning cases.
Question 167
An architecture review finds one zone outage would stop all agent controllers. What availability improvement directly addresses that?
- Plan resilient grounding-data storage and test restoration for the required failure scope
- Use separate appropriately governed projects and resource scopes for environments
- Enforce a common region tag while retaining every controller replica in the same availability zone.
- Distribute critical stateless controllers across fault-isolated instances
Explanation
Correct Answer
D:Stateless controller instances placed across fault-isolated zones reduce the risk that one zone loss stops all requests. Verify that load balancing, session state and dependent services also survive the failed zone; multiple replicas in a single zone do not meet the requirement.
Incorrect Answers
A:Redundant controllers cannot answer grounded questions if their entire document store disappears; durability, recovery and data-location requirements must be designed for the data dependency too. That is relevant to maintain usable grounding data after a storage failure, but the present scenario requires an ability to withstand a single availability-zone failure.
B:Environment separation limits accidental cross-environment changes and supports distinct access and cost controls. This helps when a team needs to isolate development and production resources; it does not establish how to withstand a single availability-zone failure.
C:Region tags do not provide fault isolation within the region. Replicate stateless controller instances across appropriate failure domains or availability zones when supported.
Question 168
An evaluator emits inconsistent grades for the same answer over repeated runs. What should be done?
- Calibrate the rubric and compare judgments with human-reviewed anchor examples
- Change the model's response temperature until its grading appears consistent
- Assess output length and usability against a mobile-oriented answer rubric
- Treat a source-citation score as proof that the evaluator rubric is stable between identical runs.
Explanation
Correct Answer
A:Repeatability and anchored scoring improve confidence in automatic evaluation. This resolves the specific need to control variance in model-based evaluation, while a measure that helps evaluate safety overblocking and missed harms addresses a different requirement.
Incorrect Answers
B:A lower sampling setting does not independently calibrate an evaluator or guarantee repeatability. Evaluation needs stable criteria, datasets and consistency checks.
C:Response usefulness includes meeting interface constraints as well as factual correctness. This helps when a team needs to test answer concision for a specified interface; it does not establish how to control variance in model-based evaluation.
D:Grounded citations are not a measure of grader repeatability. Stabilize the rubric, calibrate against human-reviewed anchors and compare repeated judgments to detect variance.
Question 169
A service expects peak concurrent tool calls to grow tenfold during enrollment week. What capacity planning activity is needed?
- Add policy tags to agent resources and assume this automatically increases model throughput and protected API rate limits.
- Route service diagnostics and traces to managed centralized monitoring resources
- Duplicate controller instances without measuring model-inference throughput, API rate limits or downstream tool concurrency.
- Load test end-to-end model calls, tool latency and downstream concurrency limits
Explanation
Correct Answer
D:AI scalability depends on model throughput and tool dependencies, not only the web server instance count. Here the relevant capability must let the team validate full pipeline capacity under a peak. An alternative targeting enforce declarative resource governance requirements solves another problem.
Incorrect Answers
A:Tags and policy assignments do not allocate additional quota or remove downstream concurrency ceilings. Perform load tests spanning model inference, tool calls and dependent service limits.
B:External observability makes post-failure investigation possible even when application instances are replaced. The function serves to retain operational telemetry independent of app instances, whereas the actual requirement is to validate full pipeline capacity under a peak.
C:More controller replicas may simply increase pressure on a constrained inference or tool backend. Load test model calls and dependent APIs under realistic enrollment peaks, then set controlled concurrency and backoff.
Question 170
A production assistant must remain source-faithful, avoid unsafe tool behavior, and prove a prompt update generalizes. Which THREE independent tests are appropriate? Choose THREE.
- Score groundedness against source references, not only fluency
- Judge tool-call safety only by whether the final text sounds polite, without inspecting actual attempted API actions.
- Assess output length and usability against a mobile-oriented answer rubric
- Use separate representative validation data and regression holdout cases
- Rerun only the prompt-update examples used during development and treat their scores as proof that new user phrasing will generalize.
- Evaluate the agent tool trajectory and policy compliance, not just final text
Explanation
Correct Answers
A:A release gate should score whether generated claims are supported by the referenced sources, but that is only one dimension. The scenario also requires separate held-out regression evidence and tool-trajectory compliance checks to assess safety and overall reliability.
D:Evaluating only on examples used to tune a prompt creates overfitting and misleading confidence.
F:A successful final answer does not justify unauthorized intermediate actions.
Incorrect Answers
B:Output politeness does not establish that an agent called only authorized tools with valid arguments. Tool trajectory and policy checks are required to assess unsafe actions.
C:Response usefulness includes meeting interface constraints as well as factual correctness. Here, however, the stated objective is to evaluate grounded answers, safe actions and out-of-sample quality; this option focuses on test answer concision for a specified interface.
E:The development set is not an independent holdout. Representative unseen regression cases are necessary to detect prompt overfitting across new inputs.