Has anyone tested DeepSeek on natural-language-to-SQL tasks where the schema is a little messy? I’m thinking of real databases with names like status, type, or created_at appearing in several tables, rather than a clean demo schema.

Does it usually ask for clarification when a request could map to more than one column, or does it confidently choose one? I’d also like to hear whether including table relationships and a few sample rows makes a noticeable difference. This is for an internal reporting tool, so a plausible but incorrect query would be a bigger problem than a request for clarification.