SnowPro Specialty: Gen AI Question 115
Single answerCortex Fine-tuningA retail company uses a Cortex-hosted base model to generate customer-support email replies. The model is fluent, but responses often miss the company's required tone, formatting, and refund-policy phrasing. The support team has 25,000 historical prompt-response pairs that were reviewed and approved by senior agents. They want to improve output consistency for this specific task while keeping all training data and operations inside Snowflake. Which approach is the MOST appropriate?
- A
Create a Cortex fine-tuning job using the approved prompt-response pairs so the model adapts to the company-specific response style and policy language
- B
Use only longer prompt instructions at inference time, because fine-tuning in Cortex is intended primarily for adding new database objects and metadata awareness
- C
Export the approved examples to an external ML platform, train a custom embedding model there, and bring embeddings back into Snowflake for text generation
- D
Replace the text-generation model with Cortex Search, because search indexes are designed to enforce tone and response formatting in generated replies
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Fine-tuning with the company's approved prompt-response pairs. In real projects, fine-tuning is most appropriate when a team has high-quality examples for a specific, repeated task and wants to improve the model's consistency, tone, structure, and domain-specific phrasing beyond what prompt engineering alone reliably achieves. This scenario matches that pattern closely: the base model is already generally capable, but it needs adaptation to a defined support workflow and company-approved response style.
By contrast, prompt-only approaches may still be useful as a first step, but they do not offer the same task-specific adaptation when large sets of curated examples already exist. Embeddings and Cortex Search support retrieval and semantic matching use cases, not direct behavioral adaptation of a generative model. Snowflake best practices for GenAI generally emphasize choosing the right tool for the problem: retrieval tools for grounding and discovery, and fine-tuning for supervised adaptation of model outputs on well-defined tasks using labeled examples kept within the Snowflake environment.
- A. Correct.
Correct. This is the strongest fit for Cortex Fine-tuning: the company has a sizable set of reviewed prompt-response examples for a narrow, repeated generation task, and the goal is to make outputs consistently follow organization-specific tone, formatting, and policy wording. Fine-tuning is appropriate when you want to adapt a supported base model's behavior using curated supervised examples while keeping the workflow in Snowflake.
- B. Incorrect.
Incorrect. Prompt engineering can help, but this option is wrong because it dismisses fine-tuning for a use case where it is well suited. Cortex fine-tuning is not primarily about adding database objects or metadata awareness. The scenario is about repeated generation behavior and stylistic/policy adherence, which is exactly where fine-tuning can be valuable when high-quality labeled examples already exist.
- C. Incorrect.
Incorrect. Embedding models are used for representing text semantically for search, retrieval, clustering, and similarity tasks. Training or using embeddings does not directly solve the problem of generating better formatted customer-support replies. This option also violates the stated preference to keep training operations inside Snowflake when Cortex fine-tuning can address the requirement directly.
- D. Incorrect.
Incorrect. Cortex Search is designed for retrieval over indexed content, not for tuning a model's generative style, formatting, or policy phrasing. Search can help retrieve relevant knowledge for grounding, but it does not replace fine-tuning when the main issue is consistent response behavior for a recurring support task.