SnowPro Specialty: Gen AI Question 241
Single answerCOMPLETE argumentsA data engineering team is building a SQL-based feature that uses Snowflake Cortex COMPLETE to draft short customer-support replies. Product requirements are strict: responses must be concise, deterministic across repeated runs for the same prompt, and must not include speculative follow-up text after the requested answer. The team notices that outputs sometimes vary slightly and occasionally continue beyond the intended one-sentence reply. Which COMPLETE argument combination is the best fit for these requirements?
- A
Set temperature to 0, set top_p to 1, and set max_tokens to a small value appropriate for one sentence
- B
Set temperature to 1, set top_p to 0.1, and omit max_tokens so the model can decide the ideal response length
- C
Set temperature to 0.7, set max_tokens high enough to avoid truncation, and rely on the prompt alone to keep the answer short
- D
Set top_p to 0, set temperature to 0, and increase max_tokens to ensure the model finishes naturally
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use COMPLETE arguments that directly enforce the desired behavior: low randomness and a bounded response length. In practice, temperature is the primary control for creativity versus determinism; setting it to 0 is the standard choice when teams want repeatable outputs for the same prompt. max_tokens is the direct argument for limiting response length, so it should be sized to the expected reply rather than left unconstrained. top_p is another sampling control, but a common best practice is to adjust temperature first and avoid unnecessary simultaneous tuning unless there is a specific reason. For Snowflake Cortex COMPLETE, these arguments are used to shape generation behavior at runtime, and effective production usage combines clear prompting with argument settings such as temperature and max_tokens to meet application requirements for consistency and brevity.
- A. Correct.
Correct. For Snowflake Cortex COMPLETE, lowering temperature reduces randomness and helps make outputs more deterministic. Using top_p of 1 avoids adding an additional sampling constraint that can interact unpredictably with temperature tuning. Setting max_tokens to a small value that matches the expected one-sentence response helps prevent the model from continuing beyond the desired length. This combination best aligns with the practical goal of concise, repeatable output.
- B. Incorrect.
Incorrect. A temperature of 1 increases randomness, which conflicts with the requirement for deterministic results. Although a low top_p can narrow token selection, omitting max_tokens leaves output length less controlled and can allow the model to continue beyond the intended one-sentence answer. This option mixes one potentially helpful control with settings that work against the stated requirements.
- C. Incorrect.
Incorrect. A moderate temperature such as 0.7 intentionally allows more variation, which is not ideal when the same prompt should produce consistent results. A high max_tokens value also increases the chance that the model will continue generating unnecessary text. Prompting matters, but relying on the prompt alone is weaker than combining prompt instructions with the appropriate COMPLETE arguments.
- D. Incorrect.
Incorrect. top_p is typically used as a probability mass threshold and should be set within the supported range; setting it to 0 is not the practical choice for controlled generation and may be invalid depending on implementation constraints. While temperature 0 supports determinism, increasing max_tokens conflicts with the need to avoid extra follow-up text. This option reflects a misunderstanding of how sampling and length controls should be applied together.