Benchmark-Driven Text-to-SQL and Discovery Evals

Planned · M2 — Internal Adoption + Design Partners3 / 6 (50%)

Objective

Set up a Dataface-native eval system centered on the Fivetran dbt SQL dataset for text-to-SQL generation and catalog/search discovery. Start with a cleaned SQL-only benchmark, add deterministic scoring and agent runners, and emit structured JSON/result tables that Dataface boards can analyze.

Milestone placement

This initiative is intentionally scoped to M2, not M1. M1 only needs a usable analyst workflow in the Fivetran analytics environment; it does not require text-to-SQL quality to be systematically benchmarked or hardened yet. The eval system becomes necessary in M2, when text-to-SQL and discovery quality must become reliable enough for repeated internal use and design-partner hardening.

Code location

All eval code lives under apps/evals/ with subdirectories per eval type (sql/, catalog/, agent/), shared utilities (shared/), benchmark data (data/), results (output/), and leaderboard dashboards (faces/). One CLI entry point: python -m apps.evals {sql,catalog,agent}. See the leaderboard task for the full directory layout.

Tasks

Infrastructure (build the measurement system)

Create cleaned dbt SQL benchmark artifact — P1 — Import raw dataset from cto-research, filter to standard SQL, produce cleaned + canary JSONL
Extract shared text-to-SQL generation function — P1 — Factor out generate_sql() so evals test the same code path as production
Build text-to-SQL eval runner and deterministic scorer — P1 — Hybrid deterministic + LLM-as-judge scoring, no warehouse needed
Build bounded non-one-shot text-to-SQL stack for local evals — P1 — Experimental plan/candidate/repair/select backend to benchmark against the one-shot generator
Add catalog discovery evals derived from SQL benchmark — P2 — IR evals (recall@k, MRR) for Dataface search/catalog tools

Visualization

Set up eval leaderboard dft project and dashboards — P1 — Dataface dashboards over eval JSONL via DuckDB, served via dft serve
~~Persist eval outputs for Dataface analysis and boards~~ — CANCELLED — merged into leaderboard task above

Dependency graph

cleaned benchmark (task 2)
  ↓
extract generate_sql ─────→ eval runner (task 3)
                             ↓
                 bounded non-one-shot backend
                              ↓
                         leaderboard dashboards
                              ↓
                    [experimentation initiative uses these]

Catalog discovery evals (task 4) depend on the cleaned benchmark but use a separate scorer.