SnowPro Specialty: Gen AI Question 11
Single answerCortex Fine-tuningA customer support team uses a Cortex-hosted base model to draft responses to warranty claims. The outputs are generally accurate, but they often miss the company’s required tone, formatting, and claim-resolution phrasing. The team has 15,000 historical prompt-response pairs from approved agent replies stored in Snowflake tables. They want to improve stylistic consistency for this specific task while keeping all data and operations inside Snowflake. Which approach is the MOST appropriate?
- A
Create a fine-tuning dataset from the approved prompt-response pairs in Snowflake and fine-tune a supported Cortex model for the warranty-claim response task
- B
Use Snowpark Container Services to train a new foundation model from scratch on the 15,000 support records because domain adaptation requires pretraining
- C
Increase the maximum output tokens and lower the temperature on the base model, because formatting and tone issues are primarily caused by generation randomness
- D
Convert the historical replies into vector embeddings and use similarity search alone, because retrieval fully replaces the need for task-specific model adaptation
Show answer and explanation
Correct answer: A
Explanation
The key signal in this scenario is that the base model is already generally accurate, but the business needs consistent task-specific outputs that match approved tone, structure, and phrasing. That is a classic fine-tuning use case: adapting a supported base model with supervised prompt-response examples for a narrow business task. Because the examples already reside in Snowflake and the requirement is to keep data and operations inside Snowflake, Cortex Fine-tuning is the most appropriate approach. In contrast, training from scratch is far too costly and unnecessary, inference parameter tuning does not provide durable task adaptation, and embeddings/vector retrieval help with grounding or recall rather than teaching response style. Snowflake best practices for Cortex Fine-tuning emphasize curated, high-quality prompt-response pairs for the target task and using fine-tuning when prompt engineering alone is insufficient to achieve consistent domain-specific behavior.
- A. Correct.
Correct. This is the best fit for the scenario. The team has a sizable set of approved prompt-response examples for a narrowly defined generation task, and they want better adherence to company-specific tone, structure, and phrasing while staying entirely within Snowflake. Cortex Fine-tuning is intended for this kind of supervised adaptation of supported models using task-specific examples stored in Snowflake.
- B. Incorrect.
Incorrect. Training a foundation model from scratch is unnecessary and unrealistic for 15,000 examples. Fine-tuning is the appropriate technique when a capable base model already exists and the goal is to adapt behavior for a specific downstream task. Snowpark Container Services can support custom ML workloads, but that does not make training a new foundation model the right answer here.
- C. Incorrect.
Incorrect. Adjusting inference parameters such as temperature or max tokens can influence variability and length, but it does not reliably teach a model company-specific phrasing patterns, resolution templates, or response style. This option reflects a common misconception that prompt or decoding tweaks can substitute for supervised adaptation when the issue is consistent task-specific behavior.
- D. Incorrect.
Incorrect. Embeddings and vector search can help retrieve relevant past cases or knowledge, but retrieval alone does not teach the model to consistently produce the desired tone, formatting, or approval-ready response style. Retrieval-augmented generation and fine-tuning solve different problems; in this case, the primary gap is behavioral/style adaptation, which fine-tuning addresses more directly.