SnowPro Specialty: Gen AI Question 12
Single answerCortex Fine-tuningA retail company uses a general-purpose large language model in Snowflake Cortex to generate customer support replies. The responses are fluent, but they often miss the company's required tone, product naming conventions, and refund-policy phrasing. The company has 40,000 historical prompt-response pairs from human agents stored in Snowflake tables and wants to improve output consistency while keeping data governance inside Snowflake. Which approach is the most appropriate?
- A
Use Cortex Fine-tuning with the historical prompt-response pairs so the model can learn the company's domain-specific response style and terminology.
- B
Convert the historical support conversations into embeddings and rely only on vector similarity search, because embeddings directly change the model's generation behavior.
- C
Create a larger virtual warehouse before calling the model, because warehouse size determines how well the model follows company-specific tone and policy language.
- D
Replace the training data with a list of company product names only, because fine-tuning works best when examples are reduced to keywords instead of full prompt-response pairs.
Show answer and explanation
Correct answer: A
Explanation
The key requirement is to adapt model outputs so they consistently reflect company-specific tone, terminology, and policy phrasing using existing supervised examples stored in Snowflake. That is a classic fine-tuning scenario. In Snowflake Cortex, fine-tuning is the appropriate approach when you need to specialize a base model using curated prompt-response data while keeping workflows and governance in Snowflake. By contrast, embeddings and vector search support retrieval patterns, not persistent model adaptation. Increasing warehouse size helps with data engineering throughput but does not alter model behavior. Best practice is to use high-quality, representative prompt-response pairs that match the target task and style. Candidates should distinguish clearly between fine-tuning for behavioral adaptation and retrieval techniques for contextual grounding.
- A. Correct.
Correct. Cortex Fine-tuning is designed for cases where an organization wants a model to better reflect domain-specific style, terminology, formatting, or task behavior using curated example pairs. In this scenario, the company already has a substantial set of historical prompt-response examples in Snowflake and wants governance to remain within Snowflake, making fine-tuning the best fit.
- B. Incorrect.
Incorrect. Embeddings and vector similarity search are useful for retrieval use cases, such as finding semantically similar past content and supplying it as context. However, embeddings do not by themselves modify the model's underlying behavior or teach it a persistent company tone. This option reflects a common confusion between retrieval-based augmentation and fine-tuning.
- C. Incorrect.
Incorrect. Virtual warehouse size affects SQL compute resources used for data preparation and related processing, but it does not improve an LLM's ability to internalize tone, policy language, or domain conventions. Model behavior in this case is not solved by scaling warehouse compute alone.
- D. Incorrect.
Incorrect. Fine-tuning requires representative training examples that show the desired relationship between inputs and outputs. Reducing the dataset to product-name keywords removes the response patterns, phrasing, and task structure the company wants the model to learn. This would weaken, not strengthen, the fine-tuning outcome.