Question 180
During operational handoff for this problem: A business team asks an LLM to answer revenue questions but the model frequently invents column names and joins. Which guardrail should be included?
Explanation
Correct Answer
Test generated SQL and model semantics against finance-approved calculations. This reduces recurrence of: Text-to-SQL accuracy depends on controlled semantic definitions and database permissions. The verification should demonstrate: A pilot maps representative natural-language revenue questions to valid SQL based on approved definitions.
Incorrect Answers
Review feature reproducibility and data lineage across training runs. That safeguard is aimed at ANOTHER risk: A retailer builds an ML feature pipeline by copying full source datasets to laptops every day. It is less targeted than the control for: Text-to-SQL accuracy depends on controlled semantic definitions and database permissions.
Pin availability checks to target regions in AI application deployment reviews. That safeguard is aimed at ANOTHER risk: A team treats Snowflake's managed AI inference as having exactly the same regional availability for every model. It is less targeted than the control for: Text-to-SQL accuracy depends on controlled semantic definitions and database permissions.
Test input parsing and malformed payload handling in ingestion contracts. That safeguard is aimed at ANOTHER risk: A migration tool defines a table column as VARCHAR then assumes all JSON path operators will work on it without conversion. It is less targeted than the control for: Text-to-SQL accuracy depends on controlled semantic definitions and database permissions.