Question 670 A stored help article contains text telling the assistant to ignore security rules. What is the most appropriate design decision? Limit retrieved context to relevant chunks within token budget Guard retrieved documents against prompt-injection instructions Include source identifiers and relevant excerpts with generated answers Retrieve only the sources the caller is authorized to access Explanation Correct Answer B:Guard retrieved documents against prompt-injection instructions fits because it treats database content as data rather than high-priority commands. That addresses the stated technical constraint. Incorrect Answers A:This alternative prevents excessive noise and unnecessary model cost. This does not meet the need for a method that treats database content as data rather than high-priority commands. C:This option makes claims traceable to retrieved passages. This does not meet the need for a method that treats database content as data rather than high-priority commands. D:That technique prevents RAG from bypassing document-level permissions. This does not meet the need for a method that treats database content as data rather than high-priority commands.
Question 671 A model cites an article but adds an exception that is absent from the source. What is the most appropriate next design step? Test prompt and context size limits on worst-case queries Use citation correctness tests on source-linked answers Evaluate answer faithfulness against retrieved source passages Evaluate refusal and escalation on sensitive unsupported requests Explanation Correct Answer C:Evaluate answer faithfulness against retrieved source passages is the appropriate choice: it detects claims unsupported by context even when fluent. The deployment should be tested with representative data. Incorrect Answers A:This alternative prevents truncation and silent context loss. This does not meet the need for a method that detects claims unsupported by context even when fluent. B:This option ensures a displayed reference actually supports the claim. This does not meet the need for a method that detects claims unsupported by context even when fluent. D:This option reduces high-impact unsupported guidance. This does not meet the need for a method that detects claims unsupported by context even when fluent.
Question 672 An assistant gives wrong answers because the correct article never reaches the prompt. The SQL deployment supports the named external model and REST integration. What is the most appropriate design decision? Track model and retrieval configuration versions for production answers Measure retrieval recall separately from generation quality Evaluate answer faithfulness against retrieved source passages Test prompt and context size limits on worst-case queries Explanation Correct Answer B:Measure retrieval recall separately from generation quality is the appropriate choice: it identifies whether failures originate in search or model reasoning. The deployment should be tested with representative data. Incorrect Answers A:This option enables reproducible investigations of changing assistant behavior. This does not meet the need for a method that identifies whether failures originate in search or model reasoning. C:This alternative detects claims unsupported by context even when fluent. This does not meet the need for a method that identifies whether failures originate in search or model reasoning. D:This alternative prevents truncation and silent context loss. This does not meet the need for a method that identifies whether failures originate in search or model reasoning.
Question 673 A finance policy assistant is asked to approve an exception absent from authorized records. Which solution best satisfies the stated constraint? Use citation correctness tests on source-linked answers Track model and retrieval configuration versions for production answers Evaluate answer faithfulness against retrieved source passages Evaluate refusal and escalation on sensitive unsupported requests Explanation Correct Answer D:This approach reduces high-impact unsupported guidance. Incorrect Answers A:This alternative ensures a displayed reference actually supports the claim. This does not meet the need for a method that reduces high-impact unsupported guidance. B:That technique enables reproducible investigations of changing assistant behavior. This does not meet the need for a method that reduces high-impact unsupported guidance. C:That technique detects claims unsupported by context even when fluent. This does not meet the need for a method that reduces high-impact unsupported guidance.
Question 674 A generated answer links to an unrelated policy version. What is the most appropriate design decision? Evaluate refusal and escalation on sensitive unsupported requests Evaluate answer faithfulness against retrieved source passages Measure retrieval recall separately from generation quality Use citation correctness tests on source-linked answers Explanation Correct Answer D:This choice ensures a displayed reference actually supports the claim. Incorrect Answers A:This option reduces high-impact unsupported guidance. This does not meet the need for a method that ensures a displayed reference actually supports the claim. B:That technique detects claims unsupported by context even when fluent. This does not meet the need for a method that ensures a displayed reference actually supports the claim. C:This option identifies whether failures originate in search or model reasoning. This does not meet the need for a method that ensures a displayed reference actually supports the claim.
Question 675 A SQL-to-model flow currently embeds a live API key in T-SQL text. What is the most appropriate design decision? Convert selected SQL rows to a compact JSON payload for model context Use managed identity or approved credential isolation for external model calls Set bounded timeout and retry behavior for SQL-external calls Move slow model calls outside long-running SQL transactions Explanation Correct Answer B:This approach avoids plaintext secrets in stored procedures and deployment scripts. Incorrect Answers A:This alternative structures authorized database facts in a model-consumable form. This does not meet the need for a method that avoids plaintext secrets in stored procedures and deployment scripts. C:This choice protects database resources when models are slow or unavailable. This does not meet the need for a method that avoids plaintext secrets in stored procedures and deployment scripts. D:That technique avoids holding row locks while waiting for network inference. This does not meet the need for a method that avoids plaintext secrets in stored procedures and deployment scripts.
Question 676 A support model needs order ID, status and timestamps from a read-only query. The SQL deployment supports the named external model and REST integration. What should the SQL developer implement? Validate and parse returned model JSON before downstream database writes Move slow model calls outside long-running SQL transactions Set bounded timeout and retry behavior for SQL-external calls Convert selected SQL rows to a compact JSON payload for model context Explanation Correct Answer D:Convert selected SQL rows to a compact JSON payload for model context fits because it structures authorized database facts in a model-consumable form. That addresses the stated technical constraint. Incorrect Answers A:This alternative prevents malformed or invented model fields from corrupting data. This does not meet the need for a method that structures authorized database facts in a model-consumable form. B:This alternative avoids holding row locks while waiting for network inference. This does not meet the need for a method that structures authorized database facts in a model-consumable form. C:This option protects database resources when models are slow or unavailable. This does not meet the need for a method that structures authorized database facts in a model-consumable form.
Question 677 An AI service is asked to issue a binding security waiver without supporting evidence. Which solution best satisfies the stated constraint? Evaluate refusal and escalation on sensitive unsupported requests Evaluate answer faithfulness against retrieved source passages Track model and retrieval configuration versions for production answers Measure retrieval recall separately from generation quality Explanation Correct Answer A:Evaluate refusal and escalation on sensitive unsupported requests is the appropriate choice: it reduces high-impact unsupported guidance. The deployment should be tested with representative data. Incorrect Answers B:This alternative detects claims unsupported by context even when fluent. This does not meet the need for a method that reduces high-impact unsupported guidance. C:This choice enables reproducible investigations of changing assistant behavior. This does not meet the need for a method that reduces high-impact unsupported guidance. D:This option identifies whether failures originate in search or model reasoning. This does not meet the need for a method that reduces high-impact unsupported guidance.
Question 678 A prompt workflow makes unbounded retries inside a transaction. What is the most appropriate next design step? Move slow model calls outside long-running SQL transactions Set bounded timeout and retry behavior for SQL-external calls Convert selected SQL rows to a compact JSON payload for model context Invoke an approved external REST endpoint from supported SQL using sp_invoke_external_rest_endpoint Explanation Correct Answer B:Set bounded timeout and retry behavior for SQL-external calls fits because it protects database resources when models are slow or unavailable. That addresses the stated technical constraint. Incorrect Answers A:That technique avoids holding row locks while waiting for network inference. This does not meet the need for a method that protects database resources when models are slow or unavailable. C:That technique structures authorized database facts in a model-consumable form. This does not meet the need for a method that protects database resources when models are slow or unavailable. D:This alternative integrates supported SQL workloads with external model APIs. This does not meet the need for a method that protects database resources when models are slow or unavailable.
Question 679 A product support bot needs to cite recently revised troubleshooting manuals. The SQL deployment supports the named external model and REST integration. Which solution best satisfies the stated constraint? Use retrieval-augmented generation for answers grounded in changing governed data Guard retrieved documents against prompt-injection instructions Require a no-answer fallback when retrieved evidence is insufficient Limit retrieved context to relevant chunks within token budget Explanation Correct Answer A:This implementation fits because it injects authorized current context into prompts before generation. That addresses the stated technical constraint. Incorrect Answers B:This option treats database content as data rather than high-priority commands. This does not meet the need for a method that injects authorized current context into prompts before generation. C:That technique avoids unsupported fabricated claims. This does not meet the need for a method that injects authorized current context into prompts before generation. D:This alternative prevents excessive noise and unnecessary model cost. This does not meet the need for a method that injects authorized current context into prompts before generation.
Question 680 An assistant cites a document whose passage does not contain the asserted procedure. What should the engineer choose? Use citation correctness tests on source-linked answers Test prompt and context size limits on worst-case queries Measure retrieval recall separately from generation quality Evaluate refusal and escalation on sensitive unsupported requests Explanation Correct Answer A:This approach ensures a displayed reference actually supports the claim. Incorrect Answers B:This choice prevents truncation and silent context loss. This does not meet the need for a method that ensures a displayed reference actually supports the claim. C:That technique identifies whether failures originate in search or model reasoning. This does not meet the need for a method that ensures a displayed reference actually supports the claim. D:This alternative reduces high-impact unsupported guidance. This does not meet the need for a method that ensures a displayed reference actually supports the claim.