SnowPro Specialty: Gen AI Question 96
Single answerCLASSIFY_TEXTA support operations team stores incoming customer emails in a Snowflake table and wants to route each email into one of four business-defined categories: BILLING, TECHNICAL_ISSUE, ACCOUNT_ACCESS, or OTHER. They need a SQL-based solution that uses Snowflake Cortex and minimizes post-processing logic in application code. Which approach best meets this requirement when using CLASSIFY_TEXT?
- A
Call CLASSIFY_TEXT on the email body and provide the four allowed labels in the function call so the model returns the best matching category directly in SQL.
- B
Call COMPLETE to generate a free-form summary of each email, then use a CASE expression to infer the category from keywords in the summary.
- C
Create vector embeddings for each email with EMBED_TEXT_768 and assign the category by selecting the label with the highest cosine similarity to the raw email text.
- D
Use SENTIMENT on each email first, then map negative emails to TECHNICAL_ISSUE, neutral emails to BILLING, and positive emails to ACCOUNT_ACCESS.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use CLASSIFY_TEXT directly with the business-defined labels. This matches the real-world need: classify unstructured email text into a predefined taxonomy in SQL, without building custom parsing or similarity logic outside Snowflake. In Snowflake Cortex, CLASSIFY_TEXT is intended for assigning text to categories, making it the most appropriate and maintainable option for routing scenarios like support triage. By contrast, COMPLETE is a general text-generation function and would require prompt engineering plus fragile downstream interpretation. Embeddings are useful for semantic search and related tasks but are not the simplest or most direct way to perform fixed-label classification. SENTIMENT measures polarity rather than category membership, so it does not solve the routing problem. This follows Snowflake best practice of choosing the Cortex function that matches the task directly rather than composing more complex workflows when a specialized function is available.
- A. Correct.
Correct. CLASSIFY_TEXT is designed for text classification tasks where you want to assign input text to one of a defined set of labels. In this scenario, the team already knows the target categories and wants the classification result returned directly in SQL with minimal application-side logic. Providing the allowed labels to CLASSIFY_TEXT aligns with the intended usage of the function for practical routing workflows.
- B. Incorrect.
Incorrect. COMPLETE can generate text, but using it to summarize first and then infer categories with keyword logic adds unnecessary complexity and reduces reliability. It also shifts classification behavior into brittle downstream parsing rules. This is a common misconception: using a general generation function for a task that has a dedicated classification function.
- C. Incorrect.
Incorrect. EMBED_TEXT_768 generates embeddings, which can support semantic similarity workflows, clustering, or retrieval, but it does not directly perform controlled label assignment. While an embedding-based classifier could be engineered, it requires extra design, label representations, similarity logic, and validation. That does not minimize post-processing and is not the best fit when CLASSIFY_TEXT exists specifically for this use case.
- D. Incorrect.
Incorrect. SENTIMENT detects emotional polarity, not business intent. A billing complaint, login problem, and feature request can all have similar sentiment but belong to different operational categories. Mapping sentiment values to routing categories is a flawed shortcut and reflects confusion between sentiment analysis and text classification.