Question 280 A bank requires comparing two Foundation models for short financial-risk explanation generation. What selection basis is defensible? Evaluate both on a representative versioned test set for accuracy, latency and safety Choose the model that generates the longest rationale Choose the newest model without running domain examples Choose solely the model with the longest context window Explanation Correct Answer A:Measured performance on the intended tasks matters more than brand or parameter size. Incorrect Answers B:Response length is not an evidence-based quality metric. C:Recency alone does not establish financial accuracy or suitability. D:Context capacity says little about task-specific correctness.
Question 281 A video analyzer must divide a training recording into topic chapters. What capability matters? Temporal segmentation with transcript and scene-change context An image model generating a different training video One unstructured summary of the entire file only Only a full-video transcript with no timestamps or scene segmentation Explanation Correct Answer A:Topic shifts and timestamps define meaningful video sections. Incorrect Answers B:Generation is not segmentation of the original source. C:A global summary cannot locate chapter boundaries. D:Transcript text alone does not identify the visual and temporal chapter boundaries.
Question 282 A multilingual RAG system receives questions in Arabic about English manuals. What should be evaluated? Only a general-purpose English embedding index with no cross-lingual evaluation Cross-lingual retrieval and translated-query quality against bilingual test cases A requirement that every user first translate the entire manual manually Only the app's maximum image resolution Explanation Correct Answer B:Language mismatch can reduce relevant document recall despite correct source content. Incorrect Answers A:English-centered vectors may not retrieve correct manual passages for Arabic queries. C:The solution can support bilingual search without repeated manual conversion. D:Image settings are unrelated to cross-language text retrieval.
Question 283 A generated image must use a horizontal 16:9 marketing layout. What should the client validate before inference? Generate at any size and stretch the finished image to 16:9 Supported size/aspect-ratio controls and expected output format Specify 16:9 only in prose without checking supported model controls Choose the closest supported square output and crop without review Explanation Correct Answer B:Generation parameters must match the requested banner geometry and model capabilities. Incorrect Answers A:Post-stretching changes proportions and may violate object-fidelity requirements. C:A prompt phrase cannot guarantee an API supports the intended aspect ratio. D:Cropping can remove required subjects or layout elements from the original design.
Question 284 A PDF indexer produces wrong reading order for a two-column document. What extraction choice can help? Layout-aware reading-order reconstruction before chunking Increase vector similarity top-k but keep incorrect text line ordering Concatenate OCR boxes by extraction order without column segmentation Chunk every fixed 1,000 characters before restoring column boundaries Explanation Correct Answer A:Correct sequencing is necessary to preserve paragraph meaning across columns. Incorrect Answers B:A retrieval parameter cannot repair sentence sequences corrupted before indexing. C:Raw extraction order may interleave unrelated left and right columns. D:Chunking already interleaved text preserves the wrong narrative order.
Question 285 A text analyzer flags all mentions of 'kill process' as violent intent, even in server logs. What boundary is missing? Use only punctuation frequency to classify safety Disable moderation for every server administrator Domain context distinguishing operational commands from threats to people Block all uses of the token kill regardless of context Explanation Correct Answer C:Technical terminology can be harmless within legitimate administration records. Incorrect Answers A:Sentence formatting does not resolve operational versus harmful meaning. B:An exception for one context does not warrant blanket removal. D:Keyword detection cannot capture the relevant target and intent.
Question 286 A translation agent cannot determine the source language and returns a guess as certain. What output contract helps? Have the next worker invent a plausible translated summary Return an empty result marked completed Always default to English and suppress the decision A typed unknown-language state with confidence and escalation pathway Explanation Correct Answer D:Explicit uncertainty prevents downstream workers from assuming translation succeeded. Incorrect Answers A:Fabrication cannot resolve missing source interpretation. B:Silent omission hides an unresolved processing step. C:A forced default can corrupt non-English inputs.
Question 287 An AI text extractor labels a policy date as current although it is stamped 'superseded.' What should be added? Remove the superseded stamp before prompting Document-version and effective-date rules in the extracted record Always trust the latest indexed hit regardless of version Use only the number of paragraphs as freshness evidence Explanation Correct Answer B:Temporal provenance controls which policy should apply to current questions. Incorrect Answers A:Hiding material context produces misleading classifications. C:Search ranking does not establish legal applicability. D:Length is independent of policy effective date.
Question 288 A knowledge bot must answer from scanned policy PDFs where ordinary text extraction returns blank pages. What ingestion step is needed? Treat blank extraction as proof the policy has no text Index only the PDF filenames Generate the missing policy from prior training knowledge OCR with document layout extraction before indexing Explanation Correct Answer D:Image-only documents need recognized text to supply searchable evidence. Incorrect Answers A:The parser may fail on image-only pages. B:File names do not contain the policy paragraphs needed for answers. C:Model memory cannot recreate an authoritative uploaded document.
Question 289 A scanned invoice has a date and a total that must be searchable for later audits. What pipeline provides this? A generated summary with invented totals OCR/layout extraction, typed fields and indexed provenance metadata Only storing the PDF bytes without an index Only full-text OCR without typed date and amount fields Explanation Correct Answer B:Structured values permit filtering while preserving page-level evidence. Incorrect Answers A:Unsupported values cannot support financial reconciliation. C:Unprocessed files do not expose searchable fields. D:Raw recognized text cannot reliably support structured numeric and date filters.
Question 290 A customer service team wants a controlled tool to check refund status but not create refunds. What function interface should be offered? An unrestricted search tool that logs in as finance Owner A deploy capability for changing the refund database schema A read-only status tool with validated refund ID and identity checks A general refund tool with silent create-if-missing behavior Explanation Correct Answer C:The operation should expose current status while withholding mutation rights. Incorrect Answers A:Broad access increases exposure beyond a single customer status lookup. B:Infrastructure write access is unnecessary to check transaction status. D:Hidden side effects can create payments during information queries.