LLM Knowledge for AI Roles

Design and implement a Cortex AI service that lets Snowflake users ask natural-language questions about their own tables and receive accurate, context-aware SQL queries back. The service must (1) accept an English question, optional table/column hints, and a target database, (2) use an LLM to generate a single, syntactically correct Snowflake SQL statement that answers the question, (3) guarantee that the generated SQL only references tables the caller’s role is authorized to read, (4) return the SQL in < 1 s P99 cold-start latency while supporting 1 000 concurrent requests, and (5) expose a REST endpoint and a Snowpark Python UDF so it can be invoked from worksheets, dashboards, and pipelines alike. You may use any Cortex LLM, vector store, caching layer, or metadata-driven prompt-engineering technique, but you must explain how you would handle hallucinated columns, optimize token usage, and keep the solution cost-efficient as the number of user databases grows.