Skip to content

Text-to-SQL Evaluation Taxonomy

A shared vocabulary for measuring text-to-SQL and data-agent systems honestly: how answers are graded, how systems fail, how often they refuse, and how stable they are from run to run.

It exists because the same system can rank first on one benchmark and third on another without getting any better or worse — the grading rule and the answer shape changed, not the capability. If you're comparing vendors or tuning your own pipeline, you need to know which of those you're looking at.

DOI

DOI: https://doi.org/10.5281/zenodo.22806083 Canonical site: https://agenticfabriq.github.io/text-to-sql-evaluation-taxonomy/ Maintainer: Agentic Fabriq, publisher of the open-source mnemiq engine and the Beacon grader Content & data: CC BY 4.0 · Code: Apache-2.0


Dimensions

Dimension Question File
Grading rule When is an answer counted correct? taxonomy/grading-rule.yaml
Answer shape How does the result's form differ from the reference? taxonomy/answer-shape.yaml
Failure mode Why was a wrong answer wrong? taxonomy/failure-mode.yaml
Refusal Did the system decline, and was that right? taxonomy/refusal.yaml
Stability Does the system give the same verdict on repeated runs? taxonomy/stability.yaml
Pre-execution check What was verified before SQL touched data? taxonomy/pre-execution-check.yaml
Semantic context What context did the model get about the schema? taxonomy/semantic-context.yaml
Benchmark corpus Which public question set, and what does it reward? taxonomy/benchmark-corpus.yaml

Five things this vocabulary makes visible

  1. Accuracy does not transfer across databases. Report per-schema spread, not just a mean.
  2. Exact match and got-the-facts can rank systems in opposite orders. Always state the grading rule next to the number.
  3. A score without a refusal rate is incomplete. A system that never declines looks more accurate than one that declines when it should.
  4. A single run hides instability. Report case-flip rate over ≥3 runs.
  5. Vendor-reported numbers on private sets aren't comparable to public ones. Label the corpus.

Data

data/published-comparison-2026-09.csv reproduces the headline comparison from the mnemiq launch article in machine-readable form, tagged with this taxonomy's grading-rule IDs. These are first-party results published by the maintainer; methodology is in the technical report, and the grader is open so anyone can re-run them.

How to cite

See CITATION.cff.

About the maintainer

Agentic Fabriq publishes mnemiq, an Apache-2.0 text-to-SQL engine where an LLM proposes SQL and deterministic checks decide whether it runs. The taxonomy is engine-agnostic and applies to any text-to-SQL system.