SnowPro Specialty: Gen AI Question 137
Single answerChoosing a modelA retail company is building a customer-support assistant in Snowflake that will summarize support cases and draft response suggestions for agents. The team must choose a model through Snowflake Cortex AI. Their priorities are: keeping inference cost under control for high daily volume, maintaining low latency for agent workflows, and achieving reliable instruction-following for summarization and response drafting. They do not need the strongest possible reasoning model for complex multi-step analysis. Which model-selection approach is the MOST appropriate?
- A
Choose a smaller, lower-cost instruction-tuned model first, validate quality on representative support data, and only move to a larger model if the quality is insufficient
- B
Choose the largest available model because larger models are always the best choice for production customer-support use cases
- C
Choose an embedding model because embeddings are optimized for summarization and response generation at lower cost
- D
Choose a model based only on maximum context window size, since latency and cost are usually similar across models in Cortex
Show answer and explanation
Correct answer: A
Explanation
The best answer is to begin with a smaller, instruction-tuned generative model and validate it against real support-case examples. In Snowflake Cortex AI, choosing a model should be driven by the workload: generation task type, quality needs, latency expectations, throughput, and cost. For routine summarization and draft-response generation, a smaller model may satisfy business requirements while reducing per-request cost and improving responsiveness. Larger models may be justified for harder reasoning tasks, but they should not be the default when the use case does not require them. Snowflake documentation and general GenAI best practices emphasize evaluating models empirically on representative prompts and data, rather than selecting solely by size or a single specification such as context window. Embedding models support retrieval and semantic search workflows, but they are not substitutes for text-generation models.
- A. Correct.
Correct. This reflects a best-practice approach to model selection: start with the smallest model that meets the task requirements, then evaluate quality, latency, and cost using representative prompts and data. For high-volume customer-support workloads, lower-cost and lower-latency models are often preferred if they provide adequate summarization and drafting quality. Instruction-following capability is important here, but the scenario does not require advanced reasoning that would justify defaulting to the largest model.
- B. Incorrect.
Incorrect. Larger models can improve quality for some tasks, but they generally come with higher cost and potentially higher latency. The scenario explicitly prioritizes cost control and low latency, and it does not require top-tier reasoning. A common misconception is assuming the biggest model is automatically the best production choice, when in practice model selection should be based on workload requirements and empirical evaluation.
- C. Incorrect.
Incorrect. Embedding models are designed for converting text into vector representations for tasks such as semantic search, retrieval, clustering, and similarity, not for generating summaries or drafting responses. Someone might choose this option because embeddings are efficient and useful in GenAI architectures, but they are not the right model type for text generation tasks.
- D. Incorrect.
Incorrect. Context window can matter when prompts or documents are large, but it is only one factor in model selection. Cost, latency, instruction-following quality, and task fit are also critical. The statement that latency and cost are usually similar across models is not accurate; these characteristics can vary meaningfully by model size and provider.