Question 10 A Content Understanding pipeline must produce agent-ready JSON from many mixed documents. Which TWO implementation controls help? Schema checks and per-file success/error reconciliation Silently fill unknown fields with generated plausible values Document classification and routing to appropriate configured analyzers Use one invoice-only schema for every document type Treat overall batch completion as proof every record succeeded Explanation Correct Answers A:A complete job must account for every uploaded input and field contract. C:Receipts, contracts and IDs require different extraction fields and expectations. Incorrect Answers B:Fabricated records undermine downstream factual accuracy. D:Noninvoice files may not contain mandatory invoice fields. E:Per-file errors can persist in finished batches.
Question 11 A compliance agent merges two policy summaries but drops an exception that applies to contractors. What QA check is useful? Use only a readability grade Select the shorter summary regardless of coverage Check source coverage for material exceptions and affected populations Regenerate with higher temperature without consulting sources Explanation Correct Answer C:A complete summary must retain rules that change obligations for contractors. Incorrect Answers A:Readable prose may omit decisive clauses. B:Brevity is not a correctness guarantee. D:Sampling changes do not identify missing policy conditions.
Question 12 A model returns a summary without citations despite being given source IDs. How should the application test this requirement? Assume any bibliography at the end establishes grounding Use an evaluation that checks claim-to-source association, not just citation count Generate random footnote markers automatically Count only the number of words in the summary Explanation Correct Answer B:Each material assertion should point to a supporting retrieved passage. Incorrect Answers A:A generic list may not substantiate individual statements. C:Invented references cannot pass source-support verification. D:Length cannot show whether claims have evidence.
Question 13 A travel agent plans an itinerary with three independent research subtasks. What scheduling strategy is efficient? Always process each search serially regardless of dependencies Run independent searches concurrently and join them by task identity Ask the model to invent results for any slow source Merge all prompts into one tool call without source attribution Explanation Correct Answer B:Parallelism is safe when no shared mutations or dependencies exist. Incorrect Answers A:Unnecessary serialization increases latency. C:Made-up availability cannot replace real search outcomes. D:Unstructured aggregation can obscure missing travel facts.
Question 14 A function tool returns a business-level failure with HTTP 200 and a `success:false` field. What should the agent use? A model-generated approval statement The HTTP code alone as proof the write committed A random generated success identifier The validated application status rather than HTTP transport status alone Explanation Correct Answer D:Transport success may still carry a rejected business operation. Incorrect Answers A:Text cannot override the backend result. B:Successful delivery of an error message is not transaction success. C:Invented receipts cannot fix a failed operation.
Question 15 A knowledge index returns stale employment policy passages after a new document version is published. What pipeline feature is missing? Version-aware ingestion and index refresh with effective-date filtering Adding more unrelated public web results Turning every retrieved passage into a new policy Increasing generation temperature on the existing passages Explanation Correct Answer A:Freshness metadata and reindexing identify superseded content before grounding. Incorrect Answers B:Unrelated pages do not identify the employer's effective version. C:A model cannot authorize publication of policy changes. D:Sampling cannot make outdated source text current.
Question 16 A support organization wants topic tags for messages without predefined categories. What result needs subsequent governance? Assign all topics the same fixed name Proposed topic clusters with validation and a controlled label taxonomy Guess unrelated topics when no evidence appears Treat every generated topic as an approved workflow label Explanation Correct Answer B:Exploratory topic extraction can inform but should not silently change routing categories. Incorrect Answers A:One category defeats useful discovery of distinct issues. C:Unsupported labels do not describe the customer's text. D:Unvetted labels can fragment reporting and automation.
Question 17 A model's response is grounded but it cites a search result that the user lacks permission to see. What should change? Replace the citation with a fabricated public link Authorization-aware retrieval with citations limited to accessible documents Make every confidential document publicly accessible Retain the citation and ask the user to request access later Explanation Correct Answer B:The user must be allowed to inspect the evidence referenced in the answer. Incorrect Answers A:False provenance does not repair the authorization problem. C:Removing security would violate private source requirements. D:The response may already disclose restricted document content.
Question 18 A PDF ingestion process must make visual tables and prose searchable. Which TWO capabilities belong in preprocessing? Strip all headings and page references before indexing Layout-aware OCR or digital parsing of document text and tables Index only the PDF filename Source-linked chunking and enrichment for figures or diagrams Generate missing table values without inspecting source Explanation Correct Answers B:Source format determines how accurate tabular content is recovered. D:Multimodal representations preserve relevant nontext evidence for retrieval. Incorrect Answers A:Discarding structure harms section and citation alignment. C:A file title does not contain all source facts. E:Invented figures cannot ground financial or technical answers.
Question 19 A TTS synthesis response is successful but its audio uses the wrong language voice. What should be checked? Only the text response's total character count The input speech language detector alone, without configuring the synthesis voice Requested locale, voice availability and final waveform language Only the storage account replication setting Explanation Correct Answer C:Transport success does not guarantee output matches the language requirement. Incorrect Answers A:Length cannot validate chosen speech language. B:Recognizing or detecting input language does not automatically select a matching output voice. D:Data durability is not speech voice configuration.
Question 20 A user asks for three options with short explanations but a prompt template returns pages of generic prose. What should be tuned? Explicit output count, length and example format, then test compliance Delete all constraints on response count Increase maximum output tokens further Replace the text model with a video generator Explanation Correct Answer A:Specific structure guides concision without changing the underlying factual task. Incorrect Answers B:Unbounded instructions do not guarantee three concise options. C:More generation capacity may worsen excessive verbosity. D:Video generation cannot satisfy the requested short textual comparison.