SnowPro Specialty: Gen AI Question 251
Single answerDifferent types of costs (virtual warehouse, EMBED_TEXT, Serving)A retail company is building a product-search assistant in Snowflake. Their pipeline has three steps: (1) nightly SQL transforms prepare product descriptions, (2) new and changed descriptions are converted into vectors using EMBED_TEXT, and (3) an end-user chat application calls a hosted model endpoint to generate answers from retrieved context. After the first month, the team wants to explain why costs appear in different places on the bill and identify which component is most directly responsible for each charge type. Which mapping is correct?
- A
Nightly SQL transforms -> virtual warehouse cost; EMBED_TEXT vector generation -> embedding/function usage cost; hosted model inference for chat responses -> Serving cost
- B
Nightly SQL transforms -> Serving cost; EMBED_TEXT vector generation -> virtual warehouse cost only; hosted model inference for chat responses -> storage cost
- C
Nightly SQL transforms -> cloud services cost only; EMBED_TEXT vector generation -> Serving cost; hosted model inference for chat responses -> virtual warehouse cost
- D
Nightly SQL transforms -> virtual warehouse cost; EMBED_TEXT vector generation -> storage cost; hosted model inference for chat responses -> virtual warehouse cost because all AI inference runs inside the selected warehouse
Show answer and explanation
Correct answer: A
Explanation
This question tests whether candidates can distinguish among three common GenAI-related Snowflake cost areas in a real solution. In practice, traditional data preparation and SQL transformations run on virtual warehouses, so those charges appear under warehouse compute. Embedding generation with EMBED_TEXT is tied to AI function usage for creating vectors, which is separate from ordinary warehouse billing. Finally, when applications call hosted models to produce generated outputs, those inference requests are associated with Serving charges rather than warehouse compute. A common misconception is that every GenAI step is covered by warehouse usage if a warehouse is involved somewhere in the architecture; Snowflake separates warehouse execution from AI function and model-serving charges. Candidates should align cost attribution to the actual service being consumed: warehouse for SQL/data processing, EMBED_TEXT for embedding generation, and Serving for hosted inference. This reflects Snowflake best practices for monitoring GenAI workloads by cost domain and reviewing billing/usage views by service category.
- A. Correct.
Correct. Standard SQL transforms executed by compute clusters are charged as virtual warehouse usage. EMBED_TEXT usage is associated with generating embeddings through Snowflake's AI function layer rather than being billed simply as warehouse runtime. Hosted model inference used to return generated chat answers is billed as Serving, which is separate from warehouse compute. This mapping reflects the distinct cost categories Snowflake surfaces for warehouse compute versus AI function/model-serving usage.
- B. Incorrect.
Incorrect. SQL transforms are not Serving; they are classic warehouse-driven compute. EMBED_TEXT is not best understood as virtual warehouse cost only, because the embedding function itself has its own AI-related usage charges. Hosted model inference for generating answers is not a storage charge; it is a model-serving/inference charge.
- C. Incorrect.
Incorrect. Nightly SQL transforms are not cloud services cost only; query execution primarily consumes warehouse compute. EMBED_TEXT is an embedding-generation function, not the same as hosted model Serving for response generation. Conversely, model inference for chat responses is not charged as virtual warehouse compute simply because the workflow exists in Snowflake.
- D. Incorrect.
Incorrect. The first mapping for SQL transforms is right, but the rest is wrong. EMBED_TEXT does not create a storage charge category as its primary cost driver; the main cost is the embedding function usage. Also, hosted model inference is not charged as warehouse compute merely because a warehouse may exist elsewhere in the pipeline. Snowflake separates model serving/inference from standard warehouse execution.