SnowPro Specialty: Gen AI Question 252
Single answerDifferent types of costs (virtual warehouse, EMBED_TEXT, Serving)A retail company is building a semantic product search application in Snowflake. Each night, a scheduled pipeline uses SQL to chunk newly added product descriptions and generate vectors with EMBED_TEXT. During the day, a customer-facing app sends search requests to a hosted LLM endpoint for answer generation. The team wants to understand which Snowflake cost categories are most directly affected by this design so they can forecast spend accurately. Which statement best describes the primary source of cost for each activity?
- A
Nightly chunking and orchestration consume virtual warehouse credits, EMBED_TEXT calls incur model-related embedding cost, and daytime hosted inference requests incur Serving cost.
- B
Nightly chunking and EMBED_TEXT both consume only virtual warehouse credits, while hosted inference requests are covered by standard cloud services charges.
- C
EMBED_TEXT charges are included in Serving cost, so only the nightly SQL pipeline uses virtual warehouse credits.
- D
All three activities are billed only as virtual warehouse usage because they are executed from SQL inside Snowflake.
Show answer and explanation
Correct answer: A
Explanation
To estimate spend correctly, candidates should separate traditional Snowflake compute from GenAI-specific usage. In this scenario, the nightly ETL-style work such as chunking text, joins, filtering, and orchestration runs on a virtual warehouse. The EMBED_TEXT function generates embeddings and is billed as AI/model-related usage rather than being treated purely as warehouse compute. Customer-facing inference against hosted models is associated with Serving cost. This distinction is important for forecasting because a team may see warehouse consumption remain steady while embedding generation or inference traffic increases significantly. Snowflake documentation and cost guidance for Cortex AI functions and model serving emphasize reviewing usage by service category rather than assuming all SQL-initiated work falls under warehouse credits.
- A. Correct.
Correct. This mapping reflects how Snowflake separates compute and model/inference-related charges. SQL-based data preparation and orchestration tasks run on a virtual warehouse and therefore consume warehouse credits. EMBED_TEXT is a model function used to generate embeddings and is billed separately from warehouse compute. Requests sent to a hosted model endpoint for inference are associated with Serving cost. This is the most accurate way to attribute spend across the described workflow.
- B. Incorrect.
Incorrect. The chunking/orchestration portion does use virtual warehouse compute, but EMBED_TEXT is not billed only as warehouse usage. It uses a model function with its own usage-based cost characteristics. Also, hosted inference for LLM requests is not simply standard cloud services charges; Snowflake tracks model serving separately as Serving cost.
- C. Incorrect.
Incorrect. EMBED_TEXT is not bundled into Serving cost in the way described here. Embedding generation and model serving are distinct usage areas. A candidate might choose this if they assume all AI model calls are part of one inference bucket, but Snowflake distinguishes embedding/model function usage from serving hosted inference endpoints.
- D. Incorrect.
Incorrect. Executing something from SQL does not mean every component is billed as warehouse compute. This is a common misconception. While the SQL statements themselves may run on a warehouse, AI functions and hosted model inference can introduce additional non-warehouse charges, including embedding-related charges and Serving cost.