SnowPro Specialty: Gen AI Question 150
Single answerStructured dataA retail company stores product, customer, and order data in Snowflake tables and wants to build a GenAI assistant that answers questions such as "Which product categories had the highest return rate in the last 90 days?" and "What was the average order value for loyalty members in Q1?" The team wants responses grounded in governed, up-to-date structured data and wants to minimize hallucinations. Which approach is the best fit?
- A
Use Cortex Analyst with a semantic model that maps business terms to the underlying structured tables and metrics.
- B
Export the tables to document files, chunk them, and index them in a vector search service so an LLM can answer directly from retrieved text.
- C
Train a custom language model on CSV exports of the tables so the model memorizes the latest business data and can answer without querying Snowflake.
- D
Convert the tables to unstructured JSON documents and use only semantic similarity search, because structured SQL-based access is less suitable for analytics questions.
Show answer and explanation
Correct answer: A
Explanation
For GenAI solutions over structured enterprise data, the strongest pattern is to keep the data in Snowflake tables and use a capability built for natural-language analytics over structured data. Cortex Analyst is intended for this use case: it relies on a semantic model so users can ask business questions in natural language while the system generates SQL against the governed underlying data. This helps reduce hallucinations because answers are grounded in live structured data rather than inferred from retrieved prose or model memory. In contrast, vector search and document chunking are better suited to unstructured content such as manuals, policies, and knowledge articles, not to precise computations over transactional tables. Best practice is to preserve structured data for SQL-based reasoning, define business semantics clearly, and let the GenAI layer translate user intent into governed analytical queries.
- A. Correct.
Correct. Cortex Analyst is designed for question answering over structured data in Snowflake. It uses a semantic model to describe facts, dimensions, relationships, and business-friendly terms so natural-language questions can be translated into accurate SQL against governed tables. This is the best fit for analytics-style questions about return rates, averages, time windows, and customer segments, while keeping answers grounded in current structured data rather than model memory.
- B. Incorrect.
Incorrect. Converting structured tables into document chunks and using vector retrieval is generally a weaker approach for precise analytical questions. Retrieval over text passages may help find relevant descriptions, but it is not the best mechanism for computing aggregates such as average order value or return rate from live tabular data. This option reflects the common misconception that all GenAI use cases should be solved with RAG over text.
- C. Incorrect.
Incorrect. Training or fine-tuning a model on CSV exports is not an appropriate way to keep answers current or reliable for operational analytics. Model training is not a substitute for querying governed data, and a trained model would not inherently stay synchronized with frequently changing order and return records. This option reflects the misconception that model weights should store business facts instead of using live data access.
- D. Incorrect.
Incorrect. Turning structured tables into unstructured JSON and relying only on semantic similarity search reduces the strengths of relational data. Questions involving filters, joins, time ranges, and numeric aggregations are better answered through SQL generated from a semantic understanding of the schema. This option incorrectly assumes semantic search is superior to structured access for analytical workloads.