Question 30 A service uses a generated summary as an input to a later compliance decision. What intermediate control fits? Allow downstream stages to treat every assertion as verified Use only the summary's readability score as approval Validate supported claims and retain exact source references between stages Discard citations after the first model summarizes Explanation Correct Answer C:Compaction must not turn uncertain evidence into unqualified policy facts. Incorrect Answers A:An earlier model's fluent output may contain errors. B:Style is separate from factual or legal validity. D:Later decisions cannot inspect the underlying support.
Question 31 A service uses a new embedding model but keeps an index built with the old vectors. What must be validated? Change the similarity threshold on the existing vector index Rename the vector field while leaving all previously generated embeddings unchanged Assume embeddings are interchangeable when both models output the same dimension Embedding compatibility, dimensions and index rebuild or version isolation Explanation Correct Answer D:Mixing representations from different encoders can make similarity comparisons invalid. Incorrect Answers A:Threshold changes cannot restore semantic compatibility between different vector encoders. B:A new field name does not transform the vector space or its stored values. C:Equal dimensionality does not mean two encoders use compatible semantic coordinates.
Question 32 An agent wants to search documents but the user has access to only one team folder. Where must authorization apply? After generation by hiding a paragraph in the UI Only by encrypting the final answer At retrieval with source ACL filtering before context assembly Only in a polite prompt about confidentiality Explanation Correct Answer C:Restricted passages must not enter the model prompt for unauthorized users. Incorrect Answers A:Data may remain exposed in the response payload. B:Encryption does not authorize retrieval of protected documents. D:Text instructions cannot replace enforcement at the source boundary.
Question 33 A document was modified after an extraction result was approved, but the workflow reuses the prior field values. What should it verify? Use model confidence from yesterday as proof of today's content Document hash/version and rerun or revalidate extraction for changed content Assume the prior extraction is immutable truth Discard the modified source and retain only the old record Explanation Correct Answer B:Approval of old bytes does not establish that new document fields are identical. Incorrect Answers A:A stale score cannot validate a different document revision. C:Source updates can change dates and amounts. D:That can conceal material changes in the authoritative document.
Question 34 A translated maintenance manual contains technical acronyms that must remain unchanged. What should the translator be instructed to do? Omit acronyms from the target version Replace part numbers with localized product names Expand every acronym using a creative guess Preserve approved terminology and acronyms while translating surrounding language Explanation Correct Answer D:Term constraints prevent dangerous changes to equipment-specific identifiers. Incorrect Answers A:Removing terms can make procedures ambiguous or unusable. B:Part identifiers are typically not linguistic translation targets. C:Guessed expansions can misstate important technical instructions.
Question 35 A multi-agent workflow combines code review, documentation and deployment. Who should control the final release? An untrusted pull-request comment that says deploy Any worker that finishes first A coordinator or external release policy gate with explicit approval state The documentation worker after generating its Markdown Explanation Correct Answer C:Specialist analysis does not implicitly authorize production deployment. Incorrect Answers A:Repository content may contain injected commands without permission. B:Completion order does not convey authority. D:Writing docs is not release authorization.
Question 36 A RAG app overuses huge chunks even for questions answered by one sentence. What evaluation should guide retrieval tuning? Measure answer support, chunk relevance, context size and retrieval latency Always use a single word per chunk Always use the largest possible chunk with no benchmarks Only optimize final answer punctuation Explanation Correct Answer A:Chunk-size tradeoffs affect both grounding precision and token cost. Incorrect Answers B:Over-fragmentation can strip necessary meaning and context. C:Excess context may hide relevant passages and increase cost. D:Punctuation does not determine retrieval effectiveness.
Question 37 A RAG source contains an instruction telling the agent to ignore the user's question. What trust boundary should be applied? Give the retrieved author full tool privileges Promote every source page directive to system instruction Treat retrieved text as evidence, not high-priority control instructions Stop requiring source IDs so directives appear cleaner Explanation Correct Answer C:Documents can inform an answer but cannot override the application's system policy. Incorrect Answers A:Document authors are not automatically authorized actors. B:Untrusted retrieved content can carry prompt injection attempts. D:Omitting provenance makes malicious source instructions harder to detect.
Question 38 An agent retrieves three chunks from separate files, one of which is missing permissions metadata. What should happen? Publish the passage and remove source identification Treat missing ACL as public read permission Fail closed for uncertain access and investigate metadata completeness Let the model decide whether the user seems trustworthy Explanation Correct Answer C:Unknown authorization cannot be treated as universally permitted retrieval. Incorrect Answers A:Concealment compounds a permissions failure. B:A metadata error could expose restricted data. D:Generated social judgment is not access control.
Question 39 A production AI app needs credentials without placing secrets in a container image. What Azure pattern helps? A developer's cached browser session mounted into production An anonymous network rule authorizing all tenants A production API key copied into image layers Managed identity with appropriate role permissions for supported services Explanation Correct Answer D:Workload identity eliminates hardcoded long-lived secrets when the target supports it. Incorrect Answers A:Interactive sessions are brittle and exceed service-specific needs. B:Network reachability is not a substitute for service authentication. C:Anyone with image access could potentially extract the secret.
Question 40 A RAG app must answer precise product-variant questions while retaining source evidence. Which TWO design choices fit? Return stable source passages and verify citation alignment Use structured SKU/version filters alongside semantic retrieval Use model memory as proof of current product specifications Generate a new specification sheet when retrieval fails Choose any high-similarity variant without identifier checks Explanation Correct Answers A:A result must actually support the generated product claim. B:Identity filtering prevents mixing similar but different product variants. Incorrect Answers C:Specifications may change and need a current source. D:Synthetic documents cannot substitute for manufacturer evidence. E:Vector similarity alone cannot establish exact product identity.