Text-to-SQL is quickly becoming one of the most practical applications of large language models (LLMs). The idea is appealing: write a question in plain English, and the system generates the correct SQL query.

But in practice, the results are mixed. Without structured schema information, models often: