Question 120
With automatic tool choice, a request can be answered from supplied context. What behavior should the application permit?
- Invent a tool call to satisfy logging expectations.
- A direct response without any tool invocation.
- Run all declared tools regardless of the response.
- Reject every response lacking a tool call.
Explanation
Correct Answer
B:Automatic selection allows the model to decide whether a tool is needed; a tool definition does not require its use every turn.
Incorrect Answers
A:Logging should reflect actual behavior rather than manufacturing nonexistent invocations.
C:Availability is not an instruction to execute every capability.
D:That imposes a forced-tool requirement absent from automatic selection.
Question 121
Output stops when a configured delimiter appears. Which response field helps distinguish this from output-budget truncation?
- The ordering of tool definitions.
- Only the response's total output-token count.
- Only the number of characters displayed.
- The stop reason and matched stop sequence.
Explanation
Correct Answer
D:These identify intentional delimiter termination rather than assuming that every short response exhausted its token allowance.
Incorrect Answers
A:Tool ordering does not identify the delimiter that terminated this response.
B:Different stop conditions can produce the same token count; the explicit stop metadata identifies why this generation ended.
C:Different termination causes can produce similarly short text.
Question 122
A stream starts with HTTP 200 but later emits an error event. How should the client behave?
- Handle the stream error and mark the response incomplete.
- Assume a final message already exists.
- Ignore errors after headers because status was successful.
- Parse the error as ordinary answer text.
Explanation
Correct Answer
A:Initial HTTP success does not guarantee the full streamed operation succeeds; event processing must retain failure handling.
Incorrect Answers
B:A terminated stream may never deliver the completion event needed to establish a full response.
C:That can publish partial output as complete despite a later execution failure.
D:Protocol errors require control-flow handling rather than presentation as the model’s answer.
Question 123
A stream sends a ping event between content events. What should the parser do?
- Execute a tool for every keepalive.
- Append the word ping to the generated answer.
- Handle it as a protocol keepalive without adding answer text.
- Restart the request on every ping.
Explanation
Correct Answer
C:Keepalive events carry no answer text; treating them as content can corrupt both the displayed response and structured event state.
Incorrect Answers
A:A ping is not a tool-use request and carries no authorized operation.
B:This confuses transport maintenance with content and corrupts the displayed response.
D:Normal keepalives do not indicate a failure requiring repeated model calls.
Question 124
Two streamed content blocks have different indexes. Why must their deltas remain separate?
- Indexes indicate interchangeable copies of the same text.
- Indexes indicate the order in which unrelated requests should be billed.
- The largest index always contains the correct answer.
- Each index identifies the block being incrementally constructed.
Explanation
Correct Answer
D:Separating buffers preserves distinct content and tool arguments when events from multiple blocks are processed.
Incorrect Answers
A:Different blocks can represent different types or operations, not redundant content.
B:Block indexes associate incremental events with content blocks inside a response, rather than defining accounting across requests.
C:An index indicates position, not semantic correctness or completion authority.
Question 125
A completed batch includes successful and failed requests. What should retry logic select?
- No requests because the batch itself completed.
- Every request regardless of its recorded outcome.
- Only failed requests whose failure conditions support retry.
- Only the first result in the file.
Explanation
Correct Answer
C:Batch completion does not imply universal item success. Inspecting individual outcomes avoids reprocessing completed work unnecessarily.
Incorrect Answers
A:A completed batch can still contain item-level errors needing remediation.
B:Blindly rerunning successes wastes resources and can duplicate downstream processing.
D:Result order does not identify which requests failed or whether retry is appropriate.
Question 126
A batch is canceled after some requests have completed. What should the consumer assume?
- Every request was never processed.
- The original inputs were automatically deleted from every system.
- Every request completed successfully.
- Inspect each item outcome rather than assuming cancellation erased completed results.
Explanation
Correct Answer
D:Cancellation can race with processing and does not roll back completed items. Inspect final per-request outcomes to distinguish completed, failed, and canceled work.
Incorrect Answers
A:Some items may have completed before cancellation took effect.
B:Canceling processing is not a comprehensive data-deletion operation.
C:Cancellation does not establish success for unfinished or failed requests.
Question 127
An application pins its API contract version. Does that by itself identify the exact model release used?
- No; setting an API version automatically selects the newest model.
- No; API versions only affect local SDK type annotations.
- Yes; any pinned API version selects one model automatically.
- No; API version and model identifier control different aspects.
Explanation
Correct Answer
D:The API version governs interface behavior, while the model field selects a model release. Record both when relevant.
Incorrect Answers
A:The API contract setting does not establish a newest-model selection policy. Model identity remains a separate request or deployment configuration.
B:A documented API version can select service interface behavior. It is not merely a local typing preference, even though it does not identify the model release.
C:An interface version does not replace explicit model selection.
Question 128
Small text in an image is unreadable in the supplied resolution. Which input change is most useful?
- Increase only the answer’s word count.
- Remove the image and retain its filename.
- Provide a clearer image or relevant higher-detail crop.
- Enlarge the same low-resolution image using only pixel duplication.
Explanation
Correct Answer
C:The model needs visual evidence with sufficient detail. Prompt changes cannot recover pixels absent from the supplied image.
Incorrect Answers
A:More output does not improve the resolution of the input evidence.
B:A filename carries less visual evidence than the already insufficient image.
D:Upscaling cannot recreate fine text detail absent from the original input; a better source or useful crop is needed.
Question 129
A request contains several images and separate questions about each. What helps preserve correct associations?
- Assume image filenames are always visible to the model.
- Combine all images into unreadable tiny thumbnails.
- Use clear labels and nearby text references identifying each image.
- Use the same unlabeled question repeatedly.
Explanation
Correct Answer
C:Stable labels and nearby questions help the model distinguish which image supplies evidence for each requested observation.
Incorrect Answers
A:The supplied representation may not expose local filenames as meaningful model context.
B:Aggressive shrinking can remove needed detail while adding association ambiguity.
D:Repeated unlabeled instructions make image-to-task mapping less clear.
Question 130
A scanned PDF contains text only as page images. Plain text extraction returns empty strings. What should ingestion use?
- A supported visual-document path or validated OCR extraction.
- Treat empty extracted text as proof the document is blank.
- Increase the language model’s temperature.
- Pass only the PDF’s file extension.
Explanation
Correct Answer
A:The information exists visually rather than in a text layer. Ingestion must access that representation before downstream reasoning.
Incorrect Answers
B:A missing text layer does not mean the rendered pages contain no information.
C:Sampling changes cannot supply document content that ingestion never provided.
D:The extension identifies format but contains none of the page evidence.