SnowPro Specialty: Gen AI Question 89
Single answerCOMPLETEA retail company stores product descriptions and customer review summaries in a Snowflake table. The analytics team wants to generate a short, natural-language marketing blurb for each product directly in SQL and save the results back into another table. They need a solution that is simple to operationalize inside Snowflake and does not require building a separate model pipeline. Which approach is the most appropriate?
- A
Use the COMPLETE function in a SQL statement to send a prompt built from the product data to a supported language model, then insert the returned text into the target table.
- B
Use Cortex Search because it automatically generates final marketing copy from table rows without requiring prompts.
- C
Export the product data to an external training environment, fine-tune a custom model, and re-import the model into Snowflake as the only supported way to create generated text from SQL.
- D
Create a dynamic table that replaces the need for any LLM call, because dynamic tables can natively rewrite product descriptions into marketing language.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use COMPLETE directly in SQL. Snowflake Cortex COMPLETE is intended for prompting supported LLMs to generate text within Snowflake, making it well suited for row-level content generation scenarios such as summarization, rewriting, classification with rationale, or marketing copy generation. In a practical implementation, a team could write a SQL statement that concatenates structured product fields into a prompt, invoke COMPLETE, and store the generated result in a destination table. This aligns with Snowflake best practices for keeping data processing close to the data and avoiding unnecessary external pipelines when managed model inference is sufficient. By contrast, Cortex Search is for retrieval use cases, dynamic tables are a data pipeline construct rather than a generation capability, and external fine-tuning introduces complexity that the scenario specifically does not require. Candidates should recognize COMPLETE as the native choice for in-database generative text completion workflows in Snowflake Cortex.
- A. Correct.
Correct. COMPLETE is designed for text generation tasks using supported large language models from within Snowflake SQL. In this scenario, the team can construct prompts from columns such as product description and review summary, call COMPLETE to generate a marketing blurb, and persist the output with standard SQL operations such as INSERT ... SELECT or CTAS. This matches the requirement to stay inside Snowflake and avoid building a separate model training pipeline.
- B. Incorrect.
Incorrect. Cortex Search is intended for search and retrieval over indexed content, not as a direct replacement for text generation. A candidate might choose this option because search and generation are often combined in GenAI architectures, but Cortex Search does not itself serve as the primary function for generating marketing copy from prompts in SQL.
- C. Incorrect.
Incorrect. Fine-tuning an external custom model is not the simplest or most appropriate solution for this requirement, and it is not the only supported way to generate text from SQL in Snowflake. The scenario explicitly asks for a straightforward in-platform approach without a separate model pipeline. COMPLETE already addresses that need.
- D. Incorrect.
Incorrect. Dynamic tables help materialize query results incrementally, but they do not natively perform LLM-based rewriting on their own. A dynamic table could be part of an orchestration pattern around generation, but it does not replace the need to call an LLM function such as COMPLETE for language generation.