Question 320 A monitoring system shows that most generated answers are syntactically valid JSON but violate contract totals. What evaluation layer is missing? Semantic business-rule validation in addition to JSON-schema checking More schema-required string fields alone Only a new chart of token latency Only removal of failed cases from the dashboard Explanation Correct Answer A:Correct types and fields do not prove arithmetic or domain invariants. Incorrect Answers B:A schema cannot calculate total consistency without application checks. C:Runtime speed is independent of numerical correctness. D:Excluding errors creates misleading apparent accuracy.
Question 321 A Foundry integration works for one environment but the SDK cannot locate deployment credentials in containers. What should be standardized? Environment-appropriate credential chain and endpoint configuration for the runtime Use arbitrary model prompts to request a new token Put production API keys in every Dockerfile Mount a personal browser profile into every container Explanation Correct Answer A:Container execution must use supported identity sources rather than a developer's local cache. Incorrect Answers B:A language model cannot issue legitimate Azure access tokens. C:Image layers can leak secrets and are difficult to rotate. D:Interactive browser state is fragile and exposes unrelated credentials.
Question 322 A video generator completes the job but omits the final five seconds of the storyboard. What should the application compare? Requested segment duration and scene coverage against actual output timeline Check generation status only and assume its video duration matches the brief Compare generated speech transcript to the storyboard without visual frames Inspect only the first representative frame of the output clip Explanation Correct Answer A:A successful generation call does not guarantee full storyboard completion. Incorrect Answers B:A successful API result need not contain every requested scene or intended final seconds. C:A transcript can match while the video omits a final silent or visual scene. D:An opening frame cannot verify completeness of the final segment.
Question 323 A customer support RAG app has poor answers to questions with synonyms not present in documents. What search enhancement can help? Stop collecting supporting passages to save context Semantic/vector matching and query-rewrite evaluation for domain synonyms Increase the moderation threshold with no retrieval change Index only exact uppercase tokens Explanation Correct Answer B:Meaning-based retrieval can find related passages despite different phrasing. Incorrect Answers A:Without grounding, the assistant may invent answers. C:Safety policy does not directly repair synonym matching. D:Stricter lexical matching can further reduce recall.
Question 324 A sentiment analyzer is reliable in English but poor on Arabic customer reviews. What should be evaluated? Arabic-language examples, tokenization, dialect handling and label calibration Only English benchmark accuracy Only the number of monthly tickets Only the image resolution of attached photos Explanation Correct Answer A:Cross-lingual performance must be measured in the affected language. Incorrect Answers B:Success in one language does not prove multilingual reliability. C:Volume cannot identify label quality across dialects. D:Visual pixels do not repair Arabic sentiment interpretation.
Question 325 A speech recognition output is returned with low confidence for an account number. What storage rule is prudent? Keep the unverified number separate from an authenticated account identifier Remove confidence metadata before processing Overwrite the verified account ID immediately Guess a plausible account number from nearby records Explanation Correct Answer A:Uncertain transcription must not replace a confirmed customer key. Incorrect Answers B:Downstream components lose a useful uncertainty signal. C:A recognition error could redirect the workflow to another person. D:Similarity does not establish the intended identity.
Question 326 A generative app includes harmful-content test prompts in CI. How should test outcomes be interpreted? Execute every command embedded in the adversarial prompts Publish the attack text with internal secrets as the test report Evaluate policy responses and regressions without treating the attack text as instructions Ignore all harmful examples because they're not user traffic Explanation Correct Answer C:Adversarial test inputs exercise safety boundaries rather than authorize actions. Incorrect Answers A:Tests frequently contain malicious instructions that must remain inert. B:Sensitive context should not be broadly disclosed through artifacts. D:Security validation requires challenging cases as well as typical requests.
Question 327 A team needs to process a video with scene descriptions and timestamped transcript segments for RAG. What tool family fits? Only static PDF OCR on the video bytes Only the video filename and duration Content Understanding video extraction producing segment-grounded text and metadata Only a text-to-video generator Explanation Correct Answer C:Time-aligned audiovisual representations enable semantic retrieval of specific scenes. Incorrect Answers A:Document OCR cannot reconstruct audio and scene sequences alone. B:Metadata cannot answer what happened during each scene. D:Generation creates new media rather than analyzing the supplied clip.
Question 328 A claims agent must compare coverage documents and then either request more details or propose a decision. What should its defined goal include? The ability to edit all insurer account records An instruction to approve every claim automatically A requirement to use the longest available reasoning response Explicit decision criteria, allowed evidence sources and escalation conditions Explanation Correct Answer D:A bounded goal prevents the agent from treating missing facts as authority to decide. Incorrect Answers A:Broad mutation rights are unnecessary for initial claim review. B:Unconditional approval ignores policy exclusions and missing evidence. C:Output length cannot substitute for verified eligibility rules.
Question 329 An AI design must process thousands of images offline for object labels but user interaction is not real time. What evaluation matters? Compare batch-compatible multimodal inference cost and throughput Use a public-only search index instead of inspecting images Use only speech output models because data are multimedia Optimize exclusively for sub-100-ms chat typing response Explanation Correct Answer A:Asynchronous work allows throughput-oriented scheduling without interactive latency constraints. Incorrect Answers B:Retrieval of unrelated text does not label supplied visuals. C:Image understanding, not audio synthesis, defines the task. D:Typing latency is irrelevant to the overnight image queue.
Question 330 An AI model is available in multiple deployment modes, but one mode lacks guaranteed throughput. What should the architect compare? Only whether sample answers look fluent Only whether each mode has the same model marketing name Documented capacity, latency, availability and cost guarantees per deployment option Only whether a model accepts a greeting prompt Explanation Correct Answer C:Operational requirements should determine which supported hosting mode is acceptable. Incorrect Answers A:A demo cannot establish production scale guarantees. B:Identical names do not imply equivalent throughput or billing. D:Greeting compatibility says little about SLOs under load.