Question 300 An agent is using Microsoft Foundry for sensitive document classification. Which test distinguishes prompt quality from access problems? Assume a model failure proves a missing SharePoint document Run only as a global administrator Ignore cases where restricted documents appear Separately test model classifications on authorized inputs and connector access under restricted identities Explanation Correct Answer D:A correct classifier cannot compensate for unauthorized input retrieval. Incorrect Answers A:Classification and retrieval are separate stages. B:Admin rights hide permission defects. C:Leaks require access enforcement review.
Question 301 A team wants to track whether a new Copilot Studio release increased tool-call errors. What measurement is appropriate? Remove older telemetry before comparison Trust a single successful demo turn Compare before/after tool failure rates by agent version and time window Compare only average message word counts Explanation Correct Answer C:Version-attributed telemetry supports regression analysis. Incorrect Answers A:Deleting baselines prevents trend analysis. B:One success cannot characterize errors at scale. D:Word count does not measure tool reliability.
Question 302 An external web agent and internal Teams agent use the same Foundry model prompt. What additional validation is needed? Test channel identity, accessible data and prompt inputs separately for each audience Assume the web user has an employee identity Send internal source documents to the website prompt by default Use only an answer-style comparison Explanation Correct Answer A:Model equivalence does not imply identical access boundaries. Incorrect Answers B:Public visitors may lack Entra authentication. C:Public inputs must remain restricted. D:Style does not validate channel authorization.
Question 303 A search tool retrieves a large document with irrelevant appendices. What optimizes grounding? Select top relevant indexed passages and keep the answer tied to accessible evidence Choose the first contiguous chunk even if the relevant heading appears much later Include every appendix in every prompt Discard the approved source and use memory Explanation Correct Answer A:Focused retrieval reduces noise and token costs. Incorrect Answers B:A position-only heuristic may exclude the passages needed to answer. C:Excess context can obscure relevant facts. D:Without retrieved approved passages the agent lacks grounding for the user's document question; relying solely on model memory can invent company-specific answers.
Question 304 A service owner wants to inspect tool usage without storing customer message text. Which Application Insights configuration is appropriate? Store model prompts in a public analytics worksheet Enable necessary event or dependency telemetry while limiting message-body logging Turn on full conversation details for all traffic indefinitely Disable all resource-role permissions Explanation Correct Answer B:Tool metrics can be tracked without collecting every personal utterance. Incorrect Answers A:A public sheet is not a secure telemetry repository. C:That increases retention and privacy risk. D:Monitoring access needs governance, not removal.
Question 305 A Foundry-connected prompt returns a model-name mismatch in production. What should be verified? Assume the connection will infer the correct model automatically The configured base-model name and deployment identifier match the authorized Azure resource Rename the chat conversation to match the model Switch to an anonymous external model URL Explanation Correct Answer B:Model connections can fail when identifiers differ from the deployment. Incorrect Answers A:The configured identifiers may be required explicitly. C:Conversation titles do not change deployments. D:That can violate enterprise controls.
Question 306 An agent sends tool telemetry to Application Insights and a security incident is suspected. Which practice helps forensic review? Allow ordinary users to edit centralized trace records Erase all failed tool calls to improve success rate Retain authorized redacted traces and protect their integrity and access Publicly expose operation IDs and payloads Explanation Correct Answer C:Auditable telemetry supports investigation without disclosing secrets. Incorrect Answers A:Editable logs are not reliable evidence. B:Deleting failed tool-call traces would make an incident response unable to establish what actions were attempted, where errors occurred or which users were affected. D:Uncontrolled sharing can leak sensitive data.
Question 307 A company wants a custom Foundry model for extracting invoice values but needs to validate cost. What should be tested? Choose the largest model solely by name Use a single short demo invoice Ignore token and capacity limits Compare realistic input token usage, response time and extraction accuracy across allowed models Explanation Correct Answer D:Model suitability includes quality and cost, not just availability. Incorrect Answers A:Size labels do not guarantee optimal cost or accuracy. B:A narrow example misses complex invoice variations. C:Production usage may exceed quota or budget.
Question 308 A knowledge search result has correct text but citations to a deleted Azure Blob document. What should be corrected? Present the dead link as verified evidence Update index metadata and source URL mappings, then reverify retrieval Remove permission checks to make old links accessible Tell the model to invent a storage path Explanation Correct Answer B:Citation links must refer to maintained source objects. Incorrect Answers A:A broken Azure Blob citation cannot substantiate the answer for the reader, so the source URL and indexed identity must be repaired and rechecked. C:Invalid links are not fixed by exposure. D:Guessed blob URLs can be wrong.
Question 309 A team wants an agent regression suite for its 20 most important support workflows. What should it build first? Test only that the agent is published and responds to the first greeting Use only successful response examples without expected outcome labels A representative test set with inputs, expected outcomes and evaluation criteria Select only unrelated generic conversational benchmarks from an open dataset Explanation Correct Answer C:A foundational set supplies repeatable evidence of core functionality. Incorrect Answers A:Availability of a chat window does not validate core support workflows. B:Unlabeled examples cannot be scored against business requirements. D:General benchmarks do not cover the organization's 20 workflows.
Question 310 An agent usually answers refund questions correctly but occasionally invents exceptions. Which quality dimension should be measured? Grounded factuality and policy consistency on adversarial refund cases Only response character count Only startup animation speed Only number of published channels Explanation Correct Answer A:The defect is unsupported policy content rather than UI rendering. Incorrect Answers B:Length does not determine policy accuracy. C:Visual latency cannot reveal unsupported exceptions. D:Channel count is unrelated to invented rules.