Question 280 An Application Insights dashboard has success metrics but no link between user requests and tool latency. What should be added? Correlation identifiers and appropriate dependency tracing at the supported telemetry scope An alphabetical list of agent names Use only daily aggregated dashboard counts without a per-request trace ID Remove timing fields from telemetry Explanation Correct Answer A:Linked traces associate a turn with its downstream operations. Incorrect Answers B:Names alone are not per-request correlation. C:Daily totals cannot correlate an individual turn to its tool latency. D:Missing durations make latency diagnosis impossible.
Question 281 A team wants to compare response quality after moving a prompt to a different Foundry model. What evaluation is appropriate? Disable regression tests before switching Assume newer models are always better on every task Run the same controlled test cases and compare factuality, format, cost and latency Check only how fast the model greets users Explanation Correct Answer C:Model changes can affect answer quality even with unchanged prompt text. Incorrect Answers A:Unreviewed model changes can introduce regressions. B:Release order does not prove suitability. D:A greeting is not the regulated extraction workload.
Question 282 A search index is organized by customer tenant. What must an agent query supply? A free-text tenant name supplied by any visitor Trust a tenant identifier supplied by the user without verifying the caller's membership An unrestricted query over all tenants An authorized tenant filter derived from verified caller context Explanation Correct Answer D:Tenant-scoped retrieval prevents cross-customer knowledge leakage. Incorrect Answers A:Untrusted text does not authorize private tenancy. B:A user could query another tenant's data by changing the identifier. C:Broad search can retrieve confidential records.
Question 283 An agent must use a Foundry deployment with regional data-handling restrictions. What should be planned? Use any region without checking compliance Assume all model endpoints have identical policy guarantees Log raw sensitive prompts to public diagnostics Verify model-region availability, data-flow policy and the approved connection configuration Explanation Correct Answer D:Selecting a model can affect where requests are processed. Incorrect Answers A:Region choice may violate approved processing boundaries. B:Deployment configurations and compliance differ. C:Data residency concerns are not resolved by unsafe logging.
Question 284 An Application Insights resource isn't receiving any events after connecting an agent. What should be tested first? Recreate all question bank content Return a fabricated dashboard metric Add an unrelated storage account and claim telemetry works Agent telemetry configuration, connection string and a controlled test interaction Explanation Correct Answer D:The resource must be correctly configured and receive real events. Incorrect Answers A:Question text doesn't repair a telemetry connection. B:False metrics conceal the missing telemetry. C:Storage provisioning does not verify event ingestion.
Question 285 A Copilot Studio team uses Azure AI Search with documents updated by an ETL job. What freshness acceptance check matters? Confirm index ingestion has completed and sample current documents are retrievable Ignore missing recent versions if old content still appears Display the ETL job name in answers as proof Assume ETL success means query index freshness Explanation Correct Answer A:Retrieval cannot see sources not yet indexed. Incorrect Answers B:Staleness can change business answers. C:A name is not evidence of indexed content. D:ETL completion and search indexing can be separate.
Question 286 A Foundry model prompt includes customer support transcripts for summarization. What data-minimization action is justified? Disable every privacy audit to reduce costs Select necessary redacted excerpts and enforce the model endpoint's approved data policy Send every customer's full chat history for context Store transcripts in the public prompt template Explanation Correct Answer B:Summarization should not transmit unrelated personal data. Incorrect Answers A:This removes safeguards while data still flows. C:Excess personal information is unnecessary. D:Templates should not contain real private data.
Question 287 An agent generates accurate answers but Application Insights reveals slow topic handoffs. What should the team optimize? Remove all authentication without investigating Correlate topic and tool durations before changing the actual bottleneck Switch every user to a different language Discard all traces to reduce storage cost Explanation Correct Answer B:Latency evidence should guide optimization. Incorrect Answers A:Security bypass is not a performance fix. C:Language changes won't repair handoff latency. D:Losing evidence prevents root-cause analysis.
Question 288 A model catalog lists a model unsupported for image inputs. A custom prompt needs to analyze screenshots. What should be verified? Use the existing text-only model and provide the screenshot filename as its entire input Skip image-specific acceptance tests Select a supported image-capable model deployment and validate its input types Assume every text model can parse images Explanation Correct Answer C:Multimodal prompts require compatible model capabilities. Incorrect Answers A:A filename does not give a text-only model the image pixels to analyze. B:Untested modality handling may fail at runtime. D:Some models accept text but not image pixels, so a screenshot-analysis prompt requires an image-capable deployment rather than a name-only model selection.
Question 289 An Azure AI Search document contains a prompt injection asking to expose the agent's credentials. What is the proper boundary? Treat indexed text as administrator instructions Use retrieved content only as data and keep tool permissions and secret handling independent Disable content monitoring because the source is internal Include credentials in a public citation Explanation Correct Answer B:A search result must not override secure agent instructions. Incorrect Answers A:Content indexing does not grant authority. C:Internal documents can still be malicious or compromised. D:Including Foundry or search credentials in a public answer can enable unauthorized service calls, regardless of whether the original search document was internal.
Question 290 A Copilot Studio agent serves both internal teams and external customers but shares an Azure AI Search index. Which architecture is necessary? Filter only by language and ignore confidentiality Assume external visitors inherit employee permissions Use one full-access query for every channel Apply channel-appropriate caller authorization and source filters before retrieval Explanation Correct Answer D:Shared indexes must not mix public and private results. Incorrect Answers A:Language is unrelated to data sensitivity. B:No such automatic trust relationship exists. C:External users could see internal content.