agentic-duckdb-analyst

A natural-language-to-SQL agent over DuckDB that doesn’t trust its own first answer.

Most text-to-SQL systems execute a query and return the result. This one runs the result through a four-tier verification ladder before surfacing it: an execution guard that catches parse and type errors for free, a result sanity check that detects empty or degenerate outputs, an intent alignment step that back-translates the SQL to verify it answers the right question, and a self-consistency check that samples independent responses and measures agreement. The system also distinguishes between answerable questions, unanswerable ones it should decline, and ambiguous prompts with multiple valid interpretations — and is evaluated on all three.

Ships with a 64-question gold eval set and reports results honestly, including known weaknesses in order-sensitive comparison and schema retrieval.