SnowPro Specialty: Gen AI Question 90
Single answerCOMPLETEA retail company stores product descriptions and customer review summaries in a Snowflake table. An analyst needs to generate short, marketing-friendly taglines for each product directly in SQL and wants the model to return only the generated text for each row. The analyst also wants to minimize prompt engineering complexity and avoid building a separate application layer. Which approach best meets these requirements when using Snowflake Cortex COMPLETE?
- A
Call SNOWFLAKE.CORTEX.COMPLETE in a SQL SELECT statement, passing a prompt built from the product fields and using options that keep the response focused on concise output.
- B
Export the table to external object storage, invoke a third-party hosted LLM outside Snowflake, and reload the generated taglines back into Snowflake.
- C
Use a standard SQL string function such as CONCAT_WS to generate the final marketing tagline, because COMPLETE cannot be used row-by-row in SQL queries.
- D
Create a dynamic table that automatically converts all product descriptions into embeddings, because COMPLETE requires vectorized input before it can generate text.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use SNOWFLAKE.CORTEX.COMPLETE directly in SQL. COMPLETE is intended for text generation tasks such as summarization, rewriting, classification-style prompting, and content creation directly where the data already resides in Snowflake. In this scenario, the analyst can build a concise prompt from product description and review summary columns and return the generated text in the SELECT result set for each row. This aligns with Snowflake best practices of reducing data movement and keeping AI workflows close to governed enterprise data. By contrast, exporting data to an external model adds unnecessary complexity, standard SQL string functions do not perform generation, and embeddings are relevant to retrieval and semantic similarity use cases rather than being required input for COMPLETE. Candidates should recognize COMPLETE as the appropriate Cortex function when the goal is row-level text generation in SQL.
- A. Correct.
Correct. SNOWFLAKE.CORTEX.COMPLETE is designed to generate text from prompts directly within SQL. A common real-world pattern is to construct a prompt from columns in each row and call COMPLETE in a SELECT statement to generate outputs such as summaries, rewrites, or taglines. This satisfies the requirement to stay inside Snowflake, avoid a separate application layer, and produce row-level generated text.
- B. Incorrect.
Incorrect. While exporting data to an external LLM could work technically, it does not best meet the stated requirements. It increases operational complexity, introduces data movement, and defeats the goal of generating results directly in SQL inside Snowflake. A candidate might choose this if they assume LLM inference must happen outside the data platform, which is a common misconception Snowflake Cortex addresses.
- C. Incorrect.
Incorrect. CONCAT_WS and similar SQL string functions can only assemble existing text; they do not perform generative AI inference. This option reflects the misconception that prompt construction alone is equivalent to text generation. COMPLETE can in fact be invoked row-by-row in SQL queries to generate new text from prompts.
- D. Incorrect.
Incorrect. Embeddings are used for semantic search, retrieval, and similarity workflows, not as a prerequisite for text generation with COMPLETE. Dynamic tables and vectorization may be useful in RAG architectures, but COMPLETE does not require vectorized input to generate a short tagline from provided text.