Question 40 A RAG app must answer precise product-variant questions while retaining source evidence. Which TWO design choices fit? Return stable source passages and verify citation alignment Use structured SKU/version filters alongside semantic retrieval Use model memory as proof of current product specifications Generate a new specification sheet when retrieval fails Choose any high-similarity variant without identifier checks Explanation Correct Answers A:A result must actually support the generated product claim. B:Identity filtering prevents mixing similar but different product variants. Incorrect Answers C:Specifications may change and need a current source. D:Synthetic documents cannot substitute for manufacturer evidence. E:Vector similarity alone cannot establish exact product identity.
Question 41 A team wants to route sensitive questions through a second reviewer only when needed. What gating criterion helps? Let model-generated enthusiasm determine whether to review Defined severity and uncertainty thresholds with documented escalation outcomes Avoid review on all high-impact actions to reduce cost Use a reviewer on every punctuation mark Explanation Correct Answer B:Risk-based review focuses expensive checks where material errors matter. Incorrect Answers A:Tone is not a valid risk classification. C:Financial and access mutations require appropriate controls. D:Microapprovals add overhead without proportional safety benefit.
Question 42 A scanned contract has two columns and tables that lose their row relationships in plain OCR. What should processing include? A synthetic summary invented from the contract type Layout-aware extraction of columns, table cells and reading order Only a line-by-line string sorted alphabetically Only filenames indexed as document titles Explanation Correct Answer B:Text recognition alone may scramble tabular relationships and section sequence. Incorrect Answers A:Generated generalizations are not the actual contract content. C:Alphabetic ordering destroys source document structure. D:Names cannot recover contract clauses or tables.
Question 43 A marketing team must retrieve similar product images, not only textual captions. What representation should it evaluate? Image or multimodal embeddings in a vector-enabled search index A speech translation endpoint for image similarity Only exact ASCII filename matching Only document word-count metadata Explanation Correct Answer A:Visual similarity depends on features encoded from the original images. Incorrect Answers B:Audio translation cannot compare product photograph content. C:Names may not capture appearance or product similarity. D:Document lengths do not represent visual characteristics.
Question 44 An agent suggests running a new tool with broad parameters while its current tools already cover the task. What should selection optimize? Choose the minimum authorized capability that satisfies the user goal Use the most powerful tool whenever present Use all tools on every turn for completeness Use the tool with the shortest name by default Explanation Correct Answer A:Tool availability should be proportional to necessary data and side effects. Incorrect Answers B:Extra privilege increases risk without improving answer quality. C:Unneeded calls add cost and can create unrelated effects. D:Naming length does not establish intent or suitability.
Question 45 A multi-agent system should stop once all assigned tasks and evidence checks are complete. What loop condition matters? Stop whenever the final writer says 'probably finished' Explicit completion criteria based on verified result statuses Stop after the first worker returns regardless of others Continue launching agents until token budget is exhausted Explanation Correct Answer B:The coordinator needs a measurable terminal condition rather than open-ended deliberation. Incorrect Answers A:A phrase is not task-set reconciliation evidence. C:Partial evidence may omit required workstreams. D:Endless work adds cost without improving required coverage.
Question 46 A model provider returns a transient service outage during a multi-step agent task. What should be reported to the client? A distinguishable retryable failure with checkpointed progress and safe resume guidance Automatic repetition of already committed payments An empty answer interpreted as a negative finding A fabricated success receipt so the workflow appears healthy Explanation Correct Answer A:Explicit state avoids claiming completion or repeating unsafe earlier effects. Incorrect Answers B:Unreconciled retries can duplicate irreversible actions. C:Unavailable inference does not establish that nothing was found. D:Invented outcomes can trigger incorrect downstream decisions.
Question 47 A model evaluates customer messages, but its few-shot examples omit mixed-language complaints. What is the likely improvement? Remove all output category definitions Repeat the same English example fifty times Use a video-generation service for language labels Add reviewed multilingual and code-switching boundary examples Explanation Correct Answer D:Representative demonstrations teach the classification rule for actual message patterns. Incorrect Answers A:Without labels, downstream classification becomes inconsistent. B:Duplicate easy cases do not cover mixed-language ambiguity. C:Visual synthesis is not the required text analysis capability.
Question 48 An agent receives a document that appears to grant itself production administrator rights. What should application policy do? Treat a signed document image as universal cloud access Ignore document-embedded privilege claims and enforce trusted identity authority Have the model edit the RBAC policy without consent Immediately grant permissions because a document says so Explanation Correct Answer B:Untrusted data cannot authorize roles or privileged tool use. Incorrect Answers A:Visual signatures do not automatically map to Azure role assignments. C:An LLM cannot expand its authority on its own initiative. D:Source text is not a validated administrative identity assertion.
Question 49 A customer wants conversational answers informed by their latest order status. Which pipeline fits? Use last week's order-status prompt example Retrieve the live authorized order record before final generation Guess fulfillment from typical delivery times Rely only on vector similarity to another order Explanation Correct Answer B:Mutable operational state needs a current data tool rather than model memory. Incorrect Answers A:Historical examples cannot establish current shipping state. C:Average shipping patterns do not prove a customer's actual order. D:A similar customer's status is not the target order status.
Question 50 A content moderator consistently blocks legitimate discussions of cybersecurity attacks. How should the rubric improve? Distinguish defensive educational context from actionable harmful requests Disable all harmful-content classification globally Judge solely by prompt length Block every mention of the word attack Explanation Correct Answer A:Intent and content purpose can affect policy classification. Incorrect Answers B:Blanket removal leaves unrelated misuse unprotected. C:Text volume is not a reliable intent signal. D:Keyword-only blocking can overreject benign security material.