SQL Dialect Conversion: LLMs Make It Easy to Build, Architecture Still Matters
Imagine a reporting product that connects to both PostgreSQL and SQL Server. A user asks for orders over 100 yuan, sorted by ID in descending order, with the first 10 rows returned. The system needs to produce two SQL statements. Turning this request into two code snippets is easy for an LLM. Making the feature maintainable, and making failures explainable, takes more work. The first decision is simple: which parts belong to the model, and which conversion rules belong in deterministic code where they can be written down and tested. ...