SnowPro Specialty: Gen AI Question 178
Single answerUpdate parametersA data science team is using Snowflake Cortex AI_COMPLETE to generate product descriptions for an e-commerce site. During testing, they notice the same prompt often produces slightly different wording, which makes it difficult to compare output quality across model versions. The team wants to make responses more deterministic without changing the prompt itself. Which update to the model parameters is the BEST choice?
- A
Decrease the temperature value so the model samples less randomly
- B
Increase max_tokens so the model has more space to complete the response consistently
- C
Increase top_p to 1.0 so the model considers more candidate tokens and becomes more deterministic
- D
Add more stop sequences so the model is forced to choose the same wording each time
Show answer and explanation
Correct answer: A
Explanation
The best answer is to decrease temperature. In Snowflake Cortex generative functions, update parameters such as temperature, top_p, max_tokens, and stop sequences affect different aspects of output generation. For improving consistency across repeated runs of the same prompt, temperature is the primary parameter to tune because it directly reduces randomness in token sampling. max_tokens affects length limits, not determinism. stop sequences affect where generation ends, not how consistently it is phrased. top_p can also influence variability, but increasing it usually broadens token choice rather than narrowing it. Snowflake documentation for Cortex completion/generation parameters describes temperature as the control for randomness and creativity, making it the most appropriate parameter to update in this scenario.
- A. Correct.
Correct. Lowering temperature reduces randomness in token selection, which generally makes model outputs more repeatable for the same prompt. In Snowflake Cortex completion-style functions, temperature is a standard generation parameter used to control creativity versus determinism. When a team wants more stable outputs for evaluation or regression testing, reducing temperature is typically the most appropriate change.
- B. Incorrect.
Incorrect. Increasing max_tokens changes the maximum allowed length of the generated output, not the degree of randomness in generation. While truncation can affect final output length, it does not directly make responses more deterministic for the same prompt.
- C. Incorrect.
Incorrect. top_p controls nucleus sampling by limiting token selection to a probability mass threshold. Raising top_p to 1.0 generally allows a broader token pool, which tends to preserve or increase variability rather than reduce it. If the goal is more deterministic behavior, lowering temperature is the better first adjustment.
- D. Incorrect.
Incorrect. Stop sequences define where generation should halt if certain text appears. They are useful for formatting and bounding output, but they do not directly make the model choose the same wording throughout the response. This option reflects a common misunderstanding between controlling output termination and controlling randomness.