Into the Hopper AI-Assisted SQL Generation with Till Döhmen
11 snips
Jan 30, 2026 
Till Döhmen, AI Lead at MotherDuck with a PhD in databases, discusses AI-assisted SQL generation and DuckDB integrations. He talks about moving text-to-SQL from benchmarks to real tools. He highlights schema and documentation importance, MCP servers and skills for reliable agents, and ways to build trust and review model-generated queries.
AI Snips
Chapters
Transcript
Episode notes
Benchmarks Don't Mirror Enterprise Reality
- Early text-to-SQL research used small, well-named benchmark databases that poorly reflected enterprise schemas.
- That mismatch limited academic results' real-world usefulness.
Provide Schema And Data Context First
- Give agents schema and data access so models won't hallucinate filter values or wrong columns.
- Plug an MCP server (or Cursor/Cloud Code) into your Postgres or DuckDB to provide that context and tools.
Models Need Built-In DB Documentation
- Even frontier LLMs lack deep knowledge of many database functions and parameters despite their age.
- Built-in, machine-readable documentation inside databases would sharply improve agent reliability.