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.
Question 331 An agent loop is slow; traces show a tool takes 90% of each request's duration. What should be optimized first? Only shorten the final reply by one sentence Add another critic agent to every call Only lower model temperature The tool's latency, batching or call frequency rather than model token speed Explanation Correct Answer D:Span timing identifies the actual bottleneck before tuning generation. Incorrect Answers A:Output brevity cannot remove a slow tool stage. B:More steps can increase total latency. C:Randomness does not repair an external service delay.
Question 332 A translator must convert customer emails while preserving operational identifiers. Which TWO controls are relevant? Retain source text and check important translated commitments against it Translate every URL and product code into natural synonyms Delete the original-language record immediately Protect immutable order IDs and names during language conversion Guess any missing account number from customer name Explanation Correct Answers A:Bilingual traceability catches altered dates and obligations. D:Identifiers should not be rewritten as translatable prose. Incorrect Answers B:Changing identifiers can break source and system references. C:Source loss prevents checking translation errors. E:Inferred IDs cannot reliably identify actual orders.
Question 333 A billing pipeline uses Content Understanding and must feed a typed downstream API. What output format fits? An unlabeled paragraph mixing all invoice facts A raw OCR screenshot with no machine-readable result A validated JSON schema with named fields and required types Random field names selected by the model on each request Explanation Correct Answer C:Structured extraction supports deterministic ingestion and application checks. Incorrect Answers A:Prose is difficult to parse and validate consistently. B:Images alone cannot populate typed service fields. D:Unstable keys break integration contracts.
Question 334 A support bot translates Chinese text for a human agent and must retain source evidence. What should it output? Only the text's character count A fabricated customer quote in English Translated text paired with the original and language metadata Only an English paraphrase without the source Explanation Correct Answer C:Side-by-side content supports verification of names and ambiguous phrases. Incorrect Answers A:Length does not provide translation content or provenance. B:Invented source wording misrepresents the customer message. D:Reviewers cannot check if technical terms were altered.
Question 335 An index for technical manuals needs to support queries about diagrams. What enrichment improves search? Use only document size as the search key Discard diagrams because vector retrieval uses text Index only words recognized in diagram titles Generate grounded figure descriptions or visual embeddings with page IDs Explanation Correct Answer D:Visual structures can become retrievable without inventing diagram contents. Incorrect Answers A:File size is not a diagram feature. B:Image features can be represented in multimodal or textual descriptions. C:Important relationships may be purely graphical.
Question 336 A factory voice agent mistakes specialized machine part names in noisy speech. What customization should be evaluated? Tune the agent's text-to-speech pronunciation dictionary only Custom speech vocabulary/model adaptation with representative domain recordings Add clean single-speaker examples only, excluding plant noise Increase language model context while leaving speech recognition unchanged Explanation Correct Answer B:Domain terminology and acoustic conditions can improve recognition accuracy. Incorrect Answers A:Improving outgoing voice pronunciation does not fix recognition of incoming spoken terms. C:Clean recordings do not represent the acoustic interference causing actual failures. D:Text reasoning cannot restore domain words misdecoded before it receives the transcript.
Question 337 A multilingual customer app needs spoken dialogue including recognition and spoken responses. What service combination fits? An image generator as an audio decoder A text-only completion endpoint with no audio processing A vector database as the speech synthesis engine Speech-to-text and text-to-speech capabilities integrated with conversational reasoning Explanation Correct Answer D:The pipeline must ingest audio and return intelligible audio output. Incorrect Answers A:Still images cannot recover words from a speech waveform. B:Written answers do not decode the customer's microphone input. C:Retrieval storage does not synthesize spoken audio.
Question 338 An enterprise wants answers grounded only in its own product documentation, not public web material. How should tools be scoped? Allow the model to invent vendor documentation for missing topics Limit retrieval to approved internal indexes and permitted source collections Use a web browser account with broad admin privileges Enable every search engine and trust a generic prompt Explanation Correct Answer B:Narrow tool exposure prevents irrelevant or unauthorized external material entering context. Incorrect Answers A:Fabrication is not a reliable substitute for product manuals. C:Administrative access is not necessary for grounded knowledge answering. D:Unrestricted retrieval may mix external guesses with internal policy.
Question 339 A DevOps engineer wants to prevent drift between tested and deployed model settings. What should be attached to the release? The model's random seed as the entire release manifest Immutable configuration references for model version, deployment, safety and prompt The operator's desktop screenshot only Only the natural-language label latest Explanation Correct Answer B:Release provenance lets failures be tied to the exact effective settings. Incorrect Answers A:A seed omits deployment version, credentials and guardrails. C:A screenshot is not a machine-verifiable release configuration. D:A mutable label cannot identify exact deployed bytes or policy.
Question 340 An app needs to answer a question about changes between two images. What comparison method fits? Generate a third illustration and compare it to neither source Joint analysis of both images with image identity and difference grounding Analyze only the first image Use document word counts as a change metric Explanation Correct Answer B:The output must attribute which visible objects changed between sources. Incorrect Answers A:Synthetic pixels add no evidence about the requested difference. C:One image cannot establish changes relative to the second. D:Word counts cannot capture visual scene differences.