Question 260 A generative app's response times increased after adding an expensive long-context evaluator. What should be measured? Only the evaluator's model parameter count Only whether the evaluator uses Markdown Evaluator marginal error detection versus latency and token cost Only monthly active users Explanation Correct Answer C:Review should be deployed where benefits justify its operational overhead. Incorrect Answers A:Size alone cannot establish the value of its findings. B:Formatting does not reveal impact on correctness or latency. D:User volume does not isolate evaluation overhead per request.
Question 261 A Foundry deployment pipeline must promote a validated prompt while keeping production secrets protected. Which TWO practices fit? Version and evaluate prompt/model/safety configuration before promotion Store long-lived production API keys directly in repository YAML Treat a passing model greeting as complete integration validation Use managed identity or secret references with scoped runtime authorization Replace production settings directly without an approved rollback artifact Explanation Correct Answers A:Reproducible releases require tested settings rather than mutable aliases. D:Credential material should not be embedded into public source or artifacts. Incorrect Answers B:Build and source access could expose reusable credentials. C:One trivial response does not exercise tools and grounding. E:Unversioned changes impede safe recovery from regressions.
Question 262 A research agent returns two conflicting source values for a financial rate. What should the coordinator do? Suppress the conflict to improve readability Average the rates without regard to definitions Use the first agent's value because it arrived earlier Preserve provenance and adjudicate version/date before presenting a conclusion Explanation Correct Answer D:Conflicting observations require evidence-based reconciliation rather than arbitrary selection. Incorrect Answers A:Hidden inconsistency misleads the final reader. B:A mean may have no valid source or conceptual meaning. C:Arrival timing does not establish correctness.
Question 263 A retrieval system must handle tables with amounts and narrative footnotes. Which data preparation helps preserve both? Treat the entire report as a single unindexed blob Strip all footnotes to shorten context Extract only page headings and omit table rows Store layout-aware tables plus nearby textual context with source locations Explanation Correct Answer D:Numeric cells often require footnote constraints for correct interpretation. Incorrect Answers A:Without structured passages, retrieval cannot target relevant data. B:Removed qualifications may change the meaning of totals. C:Headings cannot answer detailed amount questions.
Question 264 A summarizer outputs excessively long answers despite a concise requirement. Which operational change is relevant? Constrain format, length and evaluation examples, then measure compliance Disable source grounding to save space Delete all response-style instructions Increase max output tokens to encourage even longer summaries Explanation Correct Answer A:Explicit target behavior is more reliable than arbitrary truncation alone. Incorrect Answers B:Citation support should not be discarded without a content strategy. C:Unconstrained generation may be less concise. D:More capacity does not solve excess content.
Question 265 A PDF contract's table spans two pages and extraction repeats a subtotal as a new line item. What should QA inspect? Only overall OCR character count Only the number of scanned pages Page-spanning table structure, headers and row continuation rules Automatically accept every repeated number as a new charge Explanation Correct Answer C:Layout errors can duplicate financial values when tables cross boundaries. Incorrect Answers A:Total characters do not identify duplicated rows. B:Page count cannot prove correct subtotal placement. D:A repeated heading or subtotal may not represent a new item.
Question 266 A production team wants an audit trail of which model deployment served each request. What should the application log? Deployment version/identifier and correlation ID alongside request metadata Log model-family totals without distinguishing deployment instances Log request time and tenant but discard the selected deployment ID Log the configured default model but not per-request routing decisions Explanation Correct Answer A:Versioned traces tie observable behavior to the effective serving resource. Incorrect Answers B:A model family can have multiple live configurations and versions serving different requests. C:Timestamps cannot always resolve which deployment served a routed request. D:Fallbacks and canaries can use a different deployed version from the default.
Question 267 A meeting assistant must summarize both voices and slides shown during a recorded presentation. What modality combination is needed? Only text-to-speech generation of a new meeting Only sentiment classification of the meeting title Speech transcription plus visual frame analysis with synchronized timestamps Only a list of file chapter markers Explanation Correct Answer C:Neither the soundtrack nor the slides alone capture the full presentation. Incorrect Answers A:Synthetic audio is not source analysis. B:Title tone cannot summarize recorded discussions. D:Bookmarks do not contain spoken points and visual content.
Question 268 A speech transcription pipeline sometimes misidentifies question marks in excited speech. How should the application handle it? Use text-to-image generation to add punctuation Evaluate punctuation restoration against utterance meaning and downstream intent Delete all punctuation and ignore intent ambiguity Treat all generated punctuation as an exact source symbol Explanation Correct Answer B:Automatically inserted punctuation can change how a command is interpreted. Incorrect Answers A:Image synthesis cannot correct transcript semantics. C:Removal can erase important question-command distinctions. D:Spoken audio does not contain literal punctuation marks.
Question 269 A knowledge agent requires a SQL account lookup and document search within a single user request. How should capabilities be exposed? Use an image model to reconstruct table rows from memory Grant the model direct database administrator access for flexibility Separate bounded API and retrieval tools with explicit scopes and result contracts Replace both sources with a single unstructured system prompt Explanation Correct Answer C:The agent can choose structured account data or cited document evidence as needed. Incorrect Answers A:Visual generation is not an authorized data integration layer. B:Unlimited privilege is unnecessary for a targeted read-only lookup. D:Hardcoded context cannot ensure current account records or documents.
Question 270 A business wants to insert new policy documents into an index every hour. What operational monitoring is essential? Only check whether the model endpoint has a 200 status Only measure generated answer length Ingestion freshness, failed-document counts and source-to-index reconciliation Only count API users Explanation Correct Answer C:A schedule alone does not guarantee that all new content is searchable. Incorrect Answers A:Model availability does not validate search index health. B:Response length does not show index update completeness. D:Traffic is independent of ingestion success.