Question 470 A generation app repeatedly produces correct data but in inconsistent order, breaking a downstream diff. What should be enforced? Stable sorting and deterministic serialization after schema validation Set the model temperature alone and skip output normalization Accept arbitrary order while comparing raw text literally Invent missing fields to force equal-length objects Explanation Correct Answer A:Canonical formatting can make equivalent records comparable across runs. Incorrect Answers B:Sampling settings cannot guarantee field order and canonical bytes. C:Text diffs may report false changes for equivalent records. D:Synthetic data are not a correct normalization method.
Question 471 A product team needs semantic similarity lookup across thousands of support tickets. Which Azure capability fits the retrieval layer? A model deployment name as a substitute for an index An agent tracing dashboard used as a ticket database A moderation API used as the search ranking engine Vector-capable Azure AI Search with indexed embeddings and filters Explanation Correct Answer D:Vector search retrieves semantically related records beyond exact keyword matches. Incorrect Answers A:Deployment configuration does not persist searchable ticket vectors. B:Observability records do not replace a searchable domain-content index. C:Classification of unsafe content cannot retrieve related historical tickets.
Question 472 An employee agent runs with a managed identity but attempts to access another department's files. What protection is needed? Use higher model temperature to discourage cross-department reads Trust the agent's department text in its prompt Enforce per-resource entitlements at the file retrieval service Disable logging and hope no issue occurs Explanation Correct Answer C:A valid service identity does not grant universal access across datasets. Incorrect Answers A:Randomness cannot enforce document permissions. B:User-provided names are not verified authorization claims. D:Removing traces does not prevent overbroad access.
Question 473 A tenant's search answers expose internal notes marked for staff only. What is the correct architectural repair? Hide notes only in CSS after the response is delivered Enforce source ACLs in retrieval and limit the agent's allowed index access Increase the number of source citations per answer Ask users not to ask about internal notes Explanation Correct Answer B:Filtering before prompt assembly prevents unauthorized notes from reaching generation. Incorrect Answers A:Client-side hiding still exposes protected data in payloads. C:More citations may disclose additional confidential passages. D:User behavior cannot substitute for an access boundary.
Question 474 A team wants reliable service-level cost forecasting from Foundry workloads. What baseline should it use? A single extreme response multiplied by every user Only a count of employees with Microsoft accounts A random price copied from another cloud vendor Measured token distribution, request mix, model pricing and peak usage Explanation Correct Answer D:Actual workload shape and service rates support credible cost estimates. Incorrect Answers A:Worst-case-only extrapolation can misrepresent typical consumption. B:User count does not quantify token intensity and media generation use. C:Unrelated rates do not model the selected Azure deployment.
Question 475 A legal RAG query returns one supported clause and one contradictory excerpt. What should the generator do? Discard the contradictory source to make the answer concise Average both statements into an invented compromise Surface the conflict with document version and source metadata instead of silently choosing Use the first excerpt because retrieval ordered it first Explanation Correct Answer C:Conflicting evidence needs adjudication before a definitive legal conclusion. Incorrect Answers A:Suppressing material evidence misleads the reviewer. B:Some legal terms cannot be mathematically reconciled. D:Rank order does not prove which policy is effective.
Question 476 A team wants to preserve generated drafts without allowing model content to approve itself for release. Which division fits? Treat passing JSON parsing as final approval Let the model print 'approved' to trigger deployment Store draft artifacts and keep publishing behind an independent application gate Use a model-created random signature as the review receipt Explanation Correct Answer C:Generation output is not a substitute for release authorization. Incorrect Answers A:Structure validation is not an authorization decision. B:Untrusted text cannot establish release permission. D:A fabricated signature provides no accountable approval trail.
Question 477 A text-to-image prompt asks for a diagram with legible specific figures. What output check remains necessary? Trust all visible labels because the prompt specified them Use only model confidence to validate chart values Verify rendered text and numerical labels independently after generation Delete source data after producing the image Explanation Correct Answer C:Image generators can produce distorted or incorrect characters and values. Incorrect Answers A:A requested label may not be reproduced accurately. B:Self-reported assurance cannot verify pixels or underlying data. D:Data is needed to check graph and label fidelity.
Question 478 A legal researcher wants extracted document text prepared for semantic search and human reading. What format is suitable? Layout-aware Markdown retaining headings, tables and citation metadata A generated replacement legal document An arbitrary one-word summary of the whole document Only numeric confidence scores with no text Explanation Correct Answer A:Structured Markdown can preserve document organization for RAG ingestion. Incorrect Answers B:Synthetic text is not an extraction of the source. C:A one-word output loses the evidence needed for clauses. D:Confidence values do not provide searchable document content.
Question 479 An extraction workflow should detect when only 85 of 100 uploaded files yielded results. What should reconciliation compare? Only whether the overall batch reports completed Input file IDs against successful, failed and pending analyzer results Only the average token length of successful files Invent output files for the fifteen missing inputs Explanation Correct Answer B:Per-file status tracking prevents silent coverage gaps. Incorrect Answers A:A batch can finish with failures or omissions. C:Output size cannot account for missing documents. D:Fabricated data are not legitimate extraction results.
Question 480 A public support assistant produces insulting replies when users submit hostile text. What protection should be tested? Input/output safety filtering with adversarial evaluation cases Increase randomness until insults become rare Use only vector search to determine safe tone Disable all content policies to reduce false refusals Explanation Correct Answer A:Both inbound attacks and generated insults should be measured against defined policy. Incorrect Answers B:Sampling does not establish a reliable toxicity boundary. C:Retrieval similarity is not a moderation decision for generated text. D:Removing safeguards makes abusive output more likely.