Question 170 A model deployment is recreated with different parameters, but the app keeps a cached old endpoint identifier. What should be tested? Configuration refresh and actual inference against the current deployment Only the old cached inference response rather than a new deployment test Only whether yesterday's deployment succeeded Only the number of search index documents Explanation Correct Answer A:Verification must use the new resource binding rather than assuming cache invalidation. Incorrect Answers B:Previous outputs do not verify the application now targets the replacement deployment. C:Old results are not evidence for a replacement deployment. D:Index counts cannot detect a stale deployment ID.
Question 171 An AI Search index has exact ID and semantic fields but returns results from another tenant. What should be corrected? Ask the LLM to ignore cross-tenant hits after retrieval Randomize tenant IDs to test different scopes Enforced tenant filters derived from authenticated identity Increase the query similarity threshold only Explanation Correct Answer C:Search relevance cannot override per-tenant data segregation. Incorrect Answers A:Unauthorized content has already entered the model context. B:Guessing identifiers risks unauthorized record access. D:High-scoring hits can still belong to another tenant.
Question 172 A bank receives French emails and needs English summaries without altering account numbers. Which workflow is appropriate? Translate every account number into a different locale value Infer account numbers from customer names after translation Translate text while preserving exact identifiers and verifying critical values Drop all source-language content before checking the summary Explanation Correct Answer C:Localization must not rewrite account references or numeric obligations. Incorrect Answers A:Identifiers are not natural-language words to be localized. B:Guessed IDs do not identify the original transaction. D:Without source context, translation errors are harder to catch.
Question 173 A model returns good answers but an evaluation pipeline scores them incorrectly because reference answers changed. What should be versioned? Pin the candidate model version but allow reference answers to change silently Record only an aggregate pass percentage with no test-set identity Evaluation dataset, rubric and accepted reference answers with timestamps Retain prompts but drop the exact approved expected labels Explanation Correct Answer C:Measurement drift must be distinguishable from actual model behavior change. Incorrect Answers A:A stable model can appear worse when its evaluation key is changed without versioning. B:A percentage cannot reproduce the input, accepted answer and scoring rubric used. D:Without the reference labels, a later reviewer cannot reconstruct why an answer was marked wrong.
Question 174 A developer wants to ground a customer-facing answer in multiple PDF chapters while limiting token use. What retrieval design helps? Summarize all documents once and never refresh them Index titles only and ignore chapter text Paste every PDF in full on every request Relevant, bounded chunks with stable section identifiers and reranking Explanation Correct Answer D:Focused passages reduce context pressure while retaining inspectable support. Incorrect Answers A:A stale universal summary may miss updated chapter details. B:A title cannot support detailed explanatory claims. C:Unbounded context increases cost and can bury the relevant clause.
Question 175 A secured RAG system must retrieve relevant policy facts without leaking restricted sources. Which TWO design choices are necessary? Invent source references when no authorized results exist Keep source ID, version and excerpt location with each returned claim Search every company's archive and hide restricted text later Filter candidate chunks using verified user permissions before prompt assembly Use similarity score as the only authorization criterion Explanation Correct Answers B:Grounded answers need traceability to the applicable source. D:Unauthorized documents must not enter the model context. Incorrect Answers A:Fabricated citations mislead users about evidence coverage. C:The model would already receive private data. E:Relevance ranking does not establish access rights.
Question 176 A reviewer agent repeatedly labels style preferences as blocking security vulnerabilities. What severity rule should be added? Automatically block publication whenever a critic responds Require ten critical issues on every patch Ignore all code and judge only sentence length Require reproducible impact and concrete evidence for each blocking finding Explanation Correct Answer D:Severity classification should reflect real failure mechanisms rather than aesthetic preference. Incorrect Answers A:A reviewer's presence is not evidence of a material issue. B:Finding quotas encourage invented or exaggerated defects. C:Security review must inspect affected behavior and context.
Question 177 A generated corporate video must be released with synchronized subtitles and compliant branding. Which TWO checks matter? Check only the opening frame and ignore later scenes Use only an automated word count of the subtitle file Inspect final frames for approved logos and required provenance marks Assume a safe prompt guarantees every frame is acceptable Verify dialogue transcript alignment to the output timeline Explanation Correct Answers C:Visual policy must be evaluated on generated pixels across the clip. E:Subtitles need correct spoken content at the right timestamps. Incorrect Answers A:Brand or safety issues can appear after the start. B:Word counts do not establish whether text agrees with dialogue or is synchronized in time. D:Generation can introduce unanticipated visual content.
Question 178 Operations want to attribute Foundry spending to separate products running the same model family. What reporting should they record? Only generated answer character counts Only a shared unlabelled subscription total Project/deployment tags, token usage and correlated cost dimensions per workload Only search index creation date Explanation Correct Answer C:Usage attribution connects billed inference volume to the responsible service. Incorrect Answers A:Characters are not the provider's complete billable token and service dimensions. B:Aggregates cannot identify which product caused spending changes. D:Creation time does not account for ongoing inference expenses.
Question 179 A document summarizer overstates a study's conclusion, replacing 'associated with' by 'caused.' What check is needed? Remove all uncertainty terms for readability Trust the longest generated summary as most precise Use stronger action verbs to make all summaries persuasive Preserve evidence strength and causal qualifiers from the source Explanation Correct Answer D:Correlation language must not be upgraded to unsupported causation. Incorrect Answers A:Uncertainty may be essential to an accurate conclusion. B:Length cannot prove causal support. C:Stylistic emphasis can change scientific meaning.
Question 180 A coordinator runs independent agents against a rate-limited research API. Which TWO orchestration controls matter? Shared concurrency and retry budgets across delegated workers Treat 429 failures as successful empty search responses Correlation IDs with result-set reconciliation Assume each worker owns the full API quota Assign results solely by arrival order Explanation Correct Answers A:Aggregate traffic must respect the backend's quota. C:Parallel completion needs explicit mapping to assigned subtasks. Incorrect Answers B:Throttling does not prove that documents are absent. D:Uncoordinated workers can overwhelm the provider. E:Fast completion does not identify the original task.