SnowPro Specialty: Gen AI Question 229
Single answerCortex LLM Playground (Public Preview)A data science team is evaluating several large language models for an internal support chatbot. They want a fast way for analysts to compare how different prompts and models respond to the same business question before engineering builds anything in SQL or code. They also want to tune generation settings interactively during this evaluation. Which Snowflake capability is the best fit for this task?
- A
Use Cortex LLM Playground in Snowsight to interactively test prompts across supported models and adjust settings such as temperature during comparison
- B
Create a dynamic table that refreshes model responses on a schedule so analysts can compare prompts without using any interactive interface
- C
Use Snowpipe Streaming to send prompts to external models and compare responses directly inside the ingestion pipeline
- D
Build a Streamlit in Snowflake app first, because Playground is intended only for production deployment and not for prompt experimentation
Show answer and explanation
Correct answer: A
Explanation
The key requirement is interactive evaluation of prompts across models before building a production solution. Cortex LLM Playground (Public Preview) in Snowsight is designed for exactly this workflow: experimenting with prompts, comparing responses from supported models, and adjusting generation parameters in a no-code interface. In contrast, dynamic tables and Snowpipe Streaming are unrelated platform features for data transformation and ingestion, respectively. Streamlit in Snowflake is valuable when a team needs a custom user experience, but it generally comes after initial experimentation rather than replacing a built-in prompt testing environment. This reflects Snowflake best practice: use the Playground for fast validation and iteration, then operationalize successful prompts through Cortex functions, SQL, or applications as needed.
- A. Correct.
Correct. Cortex LLM Playground (Public Preview) is intended for interactive experimentation in Snowsight, allowing users to try prompts against supported LLMs and adjust inference parameters during evaluation. This aligns with the scenario: rapid prompt iteration, side-by-side exploration, and model comparison before implementation in code or SQL.
- B. Incorrect.
Incorrect. Dynamic tables are for incremental transformation pipelines, not for interactive prompt engineering or ad hoc model comparison. While a team could store prompts and responses in tables for later analysis, that does not address the need for quick, interactive experimentation in a UI.
- C. Incorrect.
Incorrect. Snowpipe Streaming is used for low-latency data ingestion into Snowflake, not for prompt testing or LLM evaluation workflows. This option reflects a common misconception that any Snowflake data pipeline feature can be repurposed for GenAI experimentation.
- D. Incorrect.
Incorrect. Streamlit in Snowflake can be used to build custom applications, but it is not the best first step when the requirement is immediate, low-effort prompt and model experimentation. The statement that Playground is only for production deployment is wrong; Playground is specifically useful during evaluation and prototyping.