Question 160
The embedding model accepts only bounded input length, while some source chunks exceed the limit. What is needed?
- Require a supported-answer fallback when the grounding set is empty or insufficient
- Reduce chunk size before embedding and preserve document-to-chunk identifiers
- Interpret similarity as retrieval relevance rather than a guarantee that a passage is correct
- Record source version or retrieval timestamp alongside the cited document identifier
Explanation
Correct Answer
B:Embedding input limits are model constraints; truncating oversized passages silently may discard important evidence. In the private-document answer assistant, confirm the fix by retesting the affected behavior after reducing chunk size before embedding and preserve document-to-chunk identifiers.
Incorrect Answers
A:Although useful elsewhere in the private-document answer assistant, rAG should not manufacture facts in the absence of evidence; an explicit abstention path is safer; the present case is different and calls for the team to reduce chunk size before embedding and preserve document-to-chunk identifiers.
C:That change addresses another failure mode: embedding proximity estimates semantic relatedness, not the authenticity or correctness of a source claim; it does not address this case, for which the relevant intervention is to reduce chunk size before embedding and preserve document-to-chunk identifiers.
D:Versioned provenance helps determine which source text supported an earlier answer in its own workflow, but that is a different correction from the present requirement to reduce chunk size before embedding and preserve document-to-chunk identifiers. The reported issue would remain.
Question 161
Two runs of a creative draft yield widely different phrasing, but reproducible regression tests are the goal. What generation setting may help if supported?
- Lower the temperature for the deployed model and API
- Check the finish reason and configured maximum output-token allowance
- Set a strict supported response schema with `additionalProperties` disabled
- Clearly delimit demonstration input and output pairs from the live user data
Explanation
Correct Answer
A:Lower sampling temperature generally reduces variability where the parameter is exposed; it does not guarantee byte-identical output. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after lowering the temperature for the deployed model and API.
Incorrect Answers
B:That change addresses another failure mode: a length-related stop indicates an output budget may have truncated generation; increasing request retries alone does not resolve it; it does not address this case, for which the relevant intervention is to lower the temperature for the deployed model and API.
C:Strict structured output constrains shape and disallows undeclared properties when the model/API supports the feature in its own workflow, but that is a different correction from the present requirement to lower the temperature for the deployed model and API. The reported issue would remain.
D:Although useful elsewhere in the instruction-following chat assistant, few-shot examples can be mistaken for active content unless their boundaries and role are explicit; delimiting reduces accidental exemplar leakage; the present case is different and calls for the team to lower the temperature for the deployed model and API.
Question 162
A few-shot prompt uses only positive examples, then fails on empty and ambiguous inputs. What should the examples cover?
- Set an output-token cap appropriate to the workflow and handle truncation safely
- Repeat evaluations across a representative dataset and track output variability
- Include representative normal cases and relevant edge cases without contaminating evaluation data
- Version and test the shared application instruction template
Explanation
Correct Answer
C:Diverse examples clarify intended behavior across boundaries, not just the easiest input pattern. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after including representative normal cases and relevant edge cases without contaminating evaluation data.
Incorrect Answers
A:Although useful elsewhere in the instruction-following chat assistant, output limits bound generated length and cost exposure but should be paired with safe handling of incomplete responses; the present case is different and calls for the team to include representative normal cases and relevant edge cases without contaminating evaluation data.
B:That change addresses another failure mode: sampling controls cannot replace empirical robustness testing, especially when infrastructure and model versions evolve; it does not address this case, for which the relevant intervention is to include representative normal cases and relevant edge cases without contaminating evaluation data.
D:Prompt rules are code-like behavior controls; versioning allows reproducible evaluations and rollbacks in its own workflow, but that is a different correction from the present requirement to include representative normal cases and relevant edge cases without contaminating evaluation data. The reported issue would remain.
Question 163
A customer-service model follows user instructions to reveal hidden policy text despite an application rule against it. Where should that rule be specified?
- Lower the temperature for the deployed model and API
- Check the finish reason and configured maximum output-token allowance
- Keep the behavioral constraint in a high-priority system or developer instruction
- Set a strict supported response schema with `additionalProperties` disabled
Explanation
Correct Answer
C:Application policy should not be placed solely in lower-priority user content that can contradict or replace it. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after keeping the behavioral constraint in a high-priority system or developer instruction.
Incorrect Answers
A:That change addresses another failure mode: lower sampling temperature generally reduces variability where the parameter is exposed; it does not guarantee byte-identical output; it does not address this case, for which the relevant intervention is to keep the behavioral constraint in a high-priority system or developer instruction.
B:A length-related stop indicates an output budget may have truncated generation; increasing request retries alone does not resolve it in its own workflow, but that is a different correction from the present requirement to keep the behavioral constraint in a high-priority system or developer instruction. The reported issue would remain.
D:Although useful elsewhere in the instruction-following chat assistant, strict structured output constrains shape and disallows undeclared properties when the model/API supports the feature; the present case is different and calls for the team to keep the behavioral constraint in a high-priority system or developer instruction.
Question 164
Two development teams disagree about response style after untracked prompt edits. Which release practice would help?
- Version and test the shared application instruction template
- Include representative normal cases and relevant edge cases without contaminating evaluation data
- Use a supported schema-constrained output mechanism rather than prose-only formatting requests
- Repeat evaluations across a representative dataset and track output variability
Explanation
Correct Answer
A:Prompt rules are code-like behavior controls; versioning allows reproducible evaluations and rollbacks. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after versioning and test the shared application instruction template.
Incorrect Answers
B:Diverse examples clarify intended behavior across boundaries, not just the easiest input pattern in its own workflow, but that is a different correction from the present requirement to version and test the shared application instruction template. The reported issue would remain.
C:Although useful elsewhere in the instruction-following chat assistant, prompt wording can encourage JSON but cannot enforce the same guarantees as supported structured outputs; the present case is different and calls for the team to version and test the shared application instruction template.
D:That change addresses another failure mode: sampling controls cannot replace empirical robustness testing, especially when infrastructure and model versions evolve; it does not address this case, for which the relevant intervention is to version and test the shared application instruction template.
Question 165
A team relies on one deterministic-looking run to prove stable model quality. What should its test design include?
- Use a supported schema-constrained output mechanism rather than prose-only formatting requests
- Repeat evaluations across a representative dataset and track output variability
- Set an output-token cap appropriate to the workflow and handle truncation safely
- Include representative normal cases and relevant edge cases without contaminating evaluation data
Explanation
Correct Answer
B:Sampling controls cannot replace empirical robustness testing, especially when infrastructure and model versions evolve. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after repeating evaluations across a representative dataset and track output variability.
Incorrect Answers
A:Although useful elsewhere in the instruction-following chat assistant, prompt wording can encourage JSON but cannot enforce the same guarantees as supported structured outputs; the present case is different and calls for the team to repeat evaluations across a representative dataset and track output variability.
C:That change addresses another failure mode: output limits bound generated length and cost exposure but should be paired with safe handling of incomplete responses; it does not address this case, for which the relevant intervention is to repeat evaluations across a representative dataset and track output variability.
D:Diverse examples clarify intended behavior across boundaries, not just the easiest input pattern in its own workflow, but that is a different correction from the present requirement to repeat evaluations across a representative dataset and track output variability. The reported issue would remain.
Question 166
An engineer sends `temperature=0` to a reasoning model that rejects custom temperature values. What should the app do?
- Declare all required fields in the supported structured-output schema
- Inspect the selected model's supported sampling parameters and omit unsupported ones
- Distinguish input context-window constraints from output-token limits
- Separate example cases from held-out evaluation questions
Explanation
Correct Answer
B:Generation controls vary by model and API; a universal temperature setting is not valid for all deployments. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after inspecting the selected model's supported sampling parameters and omit unsupported ones.
Incorrect Answers
A:That change addresses another failure mode: jSON validity alone does not ensure completeness; required properties define the application's contract; it does not address this case, for which the relevant intervention is to inspect the selected model's supported sampling parameters and omit unsupported ones.
C:Although useful elsewhere in the instruction-following chat assistant, limiting generation length does not necessarily make oversized prompt input fit a model's context window; the present case is different and calls for the team to inspect the selected model's supported sampling parameters and omit unsupported ones.
D:Leaking test examples into prompt demonstrations inflates measured quality and fails to test generalization in its own workflow, but that is a different correction from the present requirement to inspect the selected model's supported sampling parameters and omit unsupported ones. The reported issue would remain.
Question 167
A completion unexpectedly stops mid-JSON although the API request was valid. Which response signal should be inspected?
- Set a strict supported response schema with `additionalProperties` disabled
- Lower the temperature for the deployed model and API
- Clearly delimit demonstration input and output pairs from the live user data
- Check the finish reason and configured maximum output-token allowance
Explanation
Correct Answer
D:A length-related stop indicates an output budget may have truncated generation; increasing request retries alone does not resolve it. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after checking the finish reason and configured maximum output-token allowance.
Incorrect Answers
A:That change addresses another failure mode: strict structured output constrains shape and disallows undeclared properties when the model/API supports the feature; it does not address this case, for which the relevant intervention is to check the finish reason and configured maximum output-token allowance.
B:Although useful elsewhere in the instruction-following chat assistant, lower sampling temperature generally reduces variability where the parameter is exposed; it does not guarantee byte-identical output; the present case is different and calls for the team to check the finish reason and configured maximum output-token allowance.
C:Few-shot examples can be mistaken for active content unless their boundaries and role are explicit; delimiting reduces accidental exemplar leakage in its own workflow, but that is a different correction from the present requirement to check the finish reason and configured maximum output-token allowance. The reported issue would remain.
Question 168
A team responsible for an instruction-following chat assistant needs to show the model representative input-output examples for an unfamiliar formatting task; the acceptance check is that new inputs resemble the structure demonstrated by the examples. Which capability should be selected?
- Set an output token limit
- Add a few-shot example to the prompt
- Use a system or developer instruction
- Adjust the temperature parameter
Explanation
Correct Answer
B:Few-shot prompting demonstrates desired transformations through examples in the model's context. For this instruction-following chat assistant, the intended evidence is that new inputs resemble the structure demonstrated by the examples. It does not update model weights or store examples permanently after the context is discarded.
Incorrect Answers
A:The generation limit caps permitted completion tokens and can control output length and resource use; it may be used to bound the number of generated output tokens from a model request in a instruction-following chat assistant, but the current requirement is instead to show the model representative input-output examples for an unfamiliar formatting task. That separate capability does not prove that new inputs resemble the structure demonstrated by the examples; It does not determine the number of documents returned by Azure AI Search, so selecting that alternative would not meet the stated acceptance check.
C:Higher-priority instructions steer model behavior across user turns subject to the API's instruction hierarchy; it may be used to establish high-priority rules for a model's role, behavior, and response boundaries in a instruction-following chat assistant, but the current requirement is instead to show the model representative input-output examples for an unfamiliar formatting task. That separate capability does not prove that new inputs resemble the structure demonstrated by the examples; They do not grant the model permissions to private APIs or guarantee compliance against every attack, so selecting that alternative would not meet the stated acceptance check.
D:Temperature controls token-sampling randomness for supported generation models and endpoints; it may be used to reduce random variation across repeated generations when the deployed API supports it in a instruction-following chat assistant, but the current requirement is instead to show the model representative input-output examples for an unfamiliar formatting task. That separate capability does not prove that new inputs resemble the structure demonstrated by the examples; It does not control request authentication or force unsupported model output formats, so selecting that alternative would not meet the stated acceptance check.
Question 169
A formatting prompt already includes examples, yet the model sometimes copies example-specific customer IDs into new output. What prompt design should change?
- Keep the behavioral constraint in a high-priority system or developer instruction
- Set a strict supported response schema with `additionalProperties` disabled
- Check the finish reason and configured maximum output-token allowance
- Clearly delimit demonstration input and output pairs from the live user data
Explanation
Correct Answer
D:Few-shot examples can be mistaken for active content unless their boundaries and role are explicit; delimiting reduces accidental exemplar leakage. In the instruction-following chat assistant, confirm the fix by retesting the affected behavior after clearlying delimit demonstration input and output pairs from the live user data.
Incorrect Answers
A:Although useful elsewhere in the instruction-following chat assistant, application policy should not be placed solely in lower-priority user content that can contradict or replace it; the present case is different and calls for the team to clearly delimit demonstration input and output pairs from the live user data.
B:Strict structured output constrains shape and disallows undeclared properties when the model/API supports the feature in its own workflow, but that is a different correction from the present requirement to clearly delimit demonstration input and output pairs from the live user data. The reported issue would remain.
C:That change addresses another failure mode: a length-related stop indicates an output budget may have truncated generation; increasing request retries alone does not resolve it; it does not address this case, for which the relevant intervention is to clearly delimit demonstration input and output pairs from the live user data.
Question 170
An instruction-following chat assistant must bound the number of generated output tokens from a model request. Reviewers will accept the implementation when responses stop within the configured completion-token budget. Which approach best fulfills that requirement?
- Set an output token limit
- Adjust the temperature parameter
- Request a supported structured response format
- Add a few-shot example to the prompt
Explanation
Correct Answer
A:The generation limit caps permitted completion tokens and can control output length and resource use. For this instruction-following chat assistant, the intended evidence is that responses stop within the configured completion-token budget. It does not determine the number of documents returned by Azure AI Search.
Incorrect Answers
B:Temperature controls token-sampling randomness for supported generation models and endpoints; it may be used to reduce random variation across repeated generations when the deployed API supports it in a instruction-following chat assistant, but the current requirement is instead to bound the number of generated output tokens from a model request. That separate capability does not prove that responses stop within the configured completion-token budget; It does not control request authentication or force unsupported model output formats, so selecting that alternative would not meet the stated acceptance check.
C:Structured-output features constrain compatible model responses to a specified schema when available; it may be used to obtain machine-parseable JSON that conforms to a declared response schema in a instruction-following chat assistant, but the current requirement is instead to bound the number of generated output tokens from a model request. That separate capability does not prove that responses stop within the configured completion-token budget; Merely lowering temperature does not enforce object properties or valid JSON syntax, so selecting that alternative would not meet the stated acceptance check.
D:Few-shot prompting demonstrates desired transformations through examples in the model's context; it may be used to show the model representative input-output examples for an unfamiliar formatting task in a instruction-following chat assistant, but the current requirement is instead to bound the number of generated output tokens from a model request. That separate capability does not prove that responses stop within the configured completion-token budget; It does not update model weights or store examples permanently after the context is discarded, so selecting that alternative would not meet the stated acceptance check.