SnowPro Specialty: Gen AI Question 141
Single answer2.2 Perform data analysis given a use case.A retail company stores customer support tickets in a Snowflake table with columns for TICKET_ID, CREATED_AT, CUSTOMER_ID, ISSUE_TEXT, and RESOLUTION_STATUS. The GenAI team wants to identify the most common complaint themes from the last 90 days before deciding whether to build a retrieval-augmented chatbot. They need an analysis approach that scales inside Snowflake, minimizes data movement, and produces structured output that analysts can aggregate. Which approach is the BEST fit?
- A
Use AI_COMPLETE on each ISSUE_TEXT with a prompt that asks the model to return a single complaint category from a controlled list in JSON format, then aggregate the returned categories in SQL.
- B
Export the last 90 days of tickets to a local notebook, manually review a sample, and infer the most common themes without using Snowflake SQL functions.
- C
Use AI_EMBED and assume the largest-magnitude embedding dimensions correspond directly to the most common complaint themes.
- D
Use a standard SQL COUNT on ISSUE_TEXT to identify the most common complaint themes, because repeated text values represent semantic similarity.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use AI_COMPLETE to classify each ticket into a controlled set of complaint categories and then aggregate those categories with SQL. For this use case, the team needs structured, analyzable output from unstructured text. Prompting the model to return one category from a predefined taxonomy in JSON format improves consistency and enables direct downstream analysis with standard SQL. This pattern is practical for analyzing support themes, sentiment buckets, escalation reasons, or other business-defined labels.
The other options reflect common but flawed approaches. Exporting data for manual analysis breaks the requirement to minimize data movement and does not scale operationally. Using embeddings can be valuable for semantic search, deduplication, clustering, or retrieval, but embeddings do not provide directly interpretable labels without further analysis. Counting exact ISSUE_TEXT values only captures literal duplicates, not the broader semantic patterns the business wants to understand.
This aligns with Snowflake best practices for AI-assisted analytics: keep processing close to the data, use SQL-native workflows where possible, and shape model output into structured formats for reliable downstream aggregation and governance. Relevant Snowflake documentation includes Cortex AISQL function guidance for tasks such as structured generation with AI_COMPLETE and analytical workflows performed directly in Snowflake.
- A. Correct.
Correct. This approach keeps analysis in Snowflake, uses an LLM to classify unstructured ticket text into a controlled taxonomy, and returns structured output that can be grouped and counted with SQL. Asking for JSON or another constrained format is a practical best practice because it makes downstream parsing and aggregation more reliable. This is well aligned to a use case where analysts need theme frequency rather than free-form summaries.
- B. Incorrect.
Incorrect. Manual review of exported data may work for ad hoc exploration, but it does not scale well, increases data movement, and reduces reproducibility. The scenario explicitly asks for an approach that scales inside Snowflake and minimizes data movement, so exporting to a local notebook is not the best fit.
- C. Incorrect.
Incorrect. Embeddings are useful for semantic similarity, clustering, and retrieval tasks, but individual embedding dimensions are not human-interpretable complaint labels. Assuming the largest-magnitude dimensions directly map to complaint themes is a common misconception. Additional clustering or classification logic would be needed to turn embeddings into actionable categories.
- D. Incorrect.
Incorrect. A simple COUNT on raw ISSUE_TEXT only finds duplicate strings, not semantically similar complaints phrased in different ways. In support data, the same issue is often described with varied wording, so exact text frequency is not a reliable method for theme analysis.