Talk to your Data Pt. 2: Limits and Performance Enhancements

Part two of a series testing MotherDuck MCP server for natural language data analysis. The authors explore hallucination edge cases, the impact of DuckDB table/column annotations, and three prompting strategies (verbose, minimal, iterative) on query accuracy. Key findings: annotations expand the agent's exploratory scope rather than dramatically improving accuracy; prompt complexity matters less than expected; non-annotated datasets cause the agent to miss less obvious columns; and ad-hoc analysis requiring deep schema knowledge remains error-prone without explicit annotations and result verification.