Question 10 A service sends a valid JSON-schema extraction request but a safety-filtered response ends with finish_reason=content_filter. The caller attempts to parse it as a complete extraction. Which behavior is safest for the downstream application? Mark all missing fields as verified true because the schema requested them. Retry forever with no backoff or review until a full answer appears. Treat content_filter as a non-successful extraction outcome and route the request through a controlled fallback or review path. Silently overwrite the source document with the partial extraction. ExplanationCorrect Answer C:A safety-filtered response may not conform to the schema; the caller must not fabricate missing fields or assume normal completion. Other Answers A:A safety refusal is not evidence about the underlying document values. B:Unbounded retries may defeat safety controls and create cost or load without improving data correctness. D:Partial or filtered output should not replace an authoritative source record.
Question 11 An insurance workflow needs a JSON result whose keys are always exactly customer_id, action and due_date; generated prose occasionally includes a preface before the object. What directly prevents unexpected leading prose in the model response? Ask the model to be creative before returning the same object. Split the output by the first period and keep everything after it. Increase the maximum output tokens and hope the JSON begins sooner. Configure a supported structured JSON-schema response rather than depending on a “JSON only” sentence in a prompt. ExplanationCorrect Answer D:A structured-response API contract constrains the result format more reliably than a prose instruction, though parsing and validation remain required. Other Answers A:Extra creative instruction increases freedom but does not prevent an unwanted preface. B:String slicing can destroy valid JSON and is not a format guarantee. C:Longer outputs do not enforce the absence of a preface.
Question 12 A healthcare triage application accepts machine-readable JSON but must abstain when a patient record lacks the decisive measurement. Which output design makes this state distinguishable from an invented value? Include an explicit needs_clarification Boolean and a nullable finding field, then validate that missing evidence triggers the abstention state. Always fill missing measurements with the population average to preserve non-null JSON. Omit the entire response when evidence is missing and return HTTP success. Ask the model to describe its internal hidden reasoning instead of adding a status field. ExplanationCorrect Answer A:An explicit missing-data path allows consumers to distinguish unknown information from a confident but unsupported clinical inference. Other Answers B:Imputing an unobserved value can misrepresent a patient-specific fact. C:An empty success payload hides the information needed by downstream workflows. D:Hidden reasoning is neither an API field contract nor reliable evidence of measured values.
Question 13 A support-ticket classifier sometimes returns urgent, URGENT and High, but its downstream queue accepts only P1, P2 and P3. Which acceptance test most directly proves the formatting defect is fixed? Verify that token usage is below a cost ceiling for each test. Run a labeled test set and verify every parsed priority is exactly P1, P2 or P3 with no extra text or values. Confirm the model responds within two seconds on one empty prompt. Check that every response contains a period at the end. ExplanationCorrect Answer B:The test checks the specific downstream contract across varied inputs instead of merely showing that a JSON parser accepts the output. Other Answers A:Cost is important but does not prove the accepted label vocabulary is respected. C:Latency for an empty input cannot establish label correctness on real tickets. D:Punctuation is unrelated to the enumerated output contract.
Question 14 A model schema includes deeply nested oneOf, allOf and $ref constructs from a full enterprise data model. The serving endpoint rejects those constructs. What design change fits Databricks structured-output support? Append the unsupported schema as a base64 string to the response_format parameter. Use a bigger max_tokens value to activate $ref support. Use a smaller flat schema with supported properties for generation, then validate and transform into the enterprise model separately. Skip all enterprise validation because a generated object is well formed. ExplanationCorrect Answer C:The documented subset does not support the complex composition keywords; decoupling generation from enterprise validation avoids unsupported constraints. Other Answers A:Encoding does not make unsupported JSON Schema composition features valid. B:Token count is independent of JSON Schema feature support. D:Syntactic structure does not ensure compliance with the downstream enterprise contract.
Question 15 A fraud-review team wants an array of objects, each containing an invoice ID and a numeric risk_score, with no extra commentary. Which request design is the most reliable starting point? Ask for a bullet list followed by a JSON summary. Use a markdown table and rely on the parser to infer typed values. Make risk_score a natural-language sentence to improve expressiveness. Request a JSON-schema structured response for the array object shape, with typed fields and no free-form preface. ExplanationCorrect Answer D:Typed structured output states the required machine interface; downstream checks can enforce score ranges and business-specific invariants. Other Answers A:A bullet list is extra commentary and breaks the strict machine-only payload requirement. B:Markdown formatting introduces parsing ambiguity and lacks typed schema control. C:Free-form prose is not suitable for the numeric score contract.
Question 16 A generated JSON answer includes correct fields but sometimes uses long invented explanatory text inside an allowed string field. A service needs brief neutral summaries. Which additional instruction and check most directly addresses the problem? Specify the desired length and evidence-only style in the prompt, and enforce application-side checks on the string before displaying it. Add maxLength to the schema and assume Databricks enforces it fully. Replace the string with an unbounded object containing arbitrary generated subfields. Reduce retrieval citations to zero so the model has less information. ExplanationCorrect Answer A:A schema constrains field types but cannot by itself guarantee conciseness, evidence quality or all length restrictions in the supported API. Other Answers B:The documented Foundation Model API does not enforce length constraints for structured outputs. C:This broadens output freedom rather than enforcing a concise display string. D:Removing evidence can worsen factual grounding without directly controlling verbosity.
Question 17 A downstream consumer expects a list of issue records, each with title and severity. Teams want to confirm the formatting contract across model upgrades. What should the release gate validate before switching models? Approve the upgrade if a single greeting prompt sounds fluent. Run schema-conformance regression inputs and check every parsed record for required fields, permitted types and unexpected values. Compare only the two models’ total parameter counts. Skip testing because the previous model complied with the same prompt. ExplanationCorrect Answer B:A regression suite exposes model/version-specific contract drift using actual parsed outputs, not a one-off visual inspection. Other Answers A:Fluency is not evidence that structured issue records satisfy the API contract. C:Model size is unrelated to guaranteed field/type conformance. D:Output behavior can change across models and versions, so contract tests must be rerun.
Question 18 A legal team has 80-page agreements and wants one short paragraph explaining each contract’s main business obligations. Which model task best matches the requested output? Sentiment classification into positive, negative or neutral. Semantic embedding generation without an output text stage. Summarization that condenses each agreement’s major obligations into a short, faithful paragraph. Language detection to identify the contract language. ExplanationCorrect Answer C:Summarization compresses source information while preserving the central obligations; it does not simply assign a category. Other Answers A:Sentiment labels cannot communicate the actual business obligations. B:Embeddings represent similarity, not the requested readable summary. D:Identifying a language does not condense the legal requirements.
Question 19 A support center receives messages about billing, shipping and account recovery. It needs exactly one routing label for each incoming message. Which NLP task is most direct? Open-ended summarization of the full message. Machine translation into the message’s original language. Extract every noun from the message as an entity list. Text classification using the three approved destination labels, with a fallback for ambiguous cases. ExplanationCorrect Answer D:The business wants a discrete routing category, not a narrative or a free-text rewrite. Other Answers A:A summary may be useful to an agent but does not deliver the queue label required for automated routing. B:Translating text does not assign the billing, shipping or recovery category. C:Entity lists do not reliably identify the intended queue when different topics use overlapping nouns.
Question 20 A global commerce catalog contains French descriptions that English-language agents need to read while preserving product measurements and part numbers. Which task directly addresses the language barrier? Machine translation from French to English with terminology and part-number preservation checks. Clustering all French descriptions by similarity. Extracting only prices and dimensions. Generating entirely new product stories without using the source. ExplanationCorrect Answer A:Translation maps text across languages, while preserving critical identifiers must be validated separately. Other Answers B:Clusters are not English-language versions of product descriptions. C:Field extraction omits the requested readable translated descriptions. D:Unanchored generation can invent dimensions and product attributes.
