SnowPro Specialty: Gen AI Question 116
Single answerCortex Fine-tuningA retail company uses Snowflake Cortex and wants a model that rewrites product descriptions into its own brand voice while preserving product facts such as size, material, and warranty terms. The team has 40,000 examples stored in a Snowflake table, where each row contains an original product description and the approved branded rewrite. They want the lowest operational overhead and want all training data to remain inside Snowflake. Which approach is the most appropriate for this use case?
- A
Use Cortex Fine-tuning with the paired prompt/completion examples in Snowflake to customize a supported model for the rewriting task.
- B
Use Cortex Search to index the approved branded rewrites and rely on semantic retrieval alone to generate rewritten descriptions in the brand voice.
- C
Export the data from Snowflake, train a custom model externally, and then call it from Snowflake because Cortex Fine-tuning requires training data to be outside Snowflake.
- D
Use a text embedding model on the original descriptions and treat the embedding vectors as the fine-tuned model output for production inference.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Cortex Fine-tuning because the problem is a classic supervised text-transformation use case: the company has many examples of desired outputs and wants a model to learn a specific style and format. This is more appropriate than retrieval-based approaches such as Cortex Search, which help provide relevant context but do not by themselves teach a model to produce a branded rewrite pattern consistently. It is also preferable to exporting data for external model training because Cortex Fine-tuning is intended to let teams customize supported models while keeping data and operations within Snowflake. In Snowflake best practices, fine-tuning is most suitable when you have high-quality labeled examples for a repeatable task and need behavior tailored beyond prompt engineering alone.
- A. Correct.
Correct. This scenario is a strong fit for Cortex Fine-tuning because the company has a large set of supervised input/output examples for a narrowly defined generation task: rewriting text into a consistent brand voice while preserving factual content. Cortex Fine-tuning is designed for adapting supported foundation models using labeled examples stored in Snowflake, which aligns with the requirement to keep data inside Snowflake and minimize operational overhead.
- B. Incorrect.
Incorrect. Cortex Search is intended for retrieval over indexed content, not for supervised adaptation of a model's writing style. Search could help fetch similar examples or policy text, but semantic retrieval alone does not fine-tune a model to consistently transform new product descriptions into a branded rewrite. Choosing this option reflects the common misconception that retrieval and fine-tuning solve the same problem.
- C. Incorrect.
Incorrect. This directly conflicts with the stated requirement for low operational overhead and keeping training data in Snowflake. Cortex Fine-tuning exists specifically to support model customization workflows in Snowflake. Exporting data to an external environment adds unnecessary complexity, governance overhead, and data movement, which is the opposite of what the team wants.
- D. Incorrect.
Incorrect. Embeddings are vector representations useful for similarity search, clustering, and retrieval-related tasks. They are not generated text outputs and cannot replace a fine-tuned generative model for rewriting descriptions. This option confuses representation learning with text generation.