Text-to-SQL Engine, Upstream Evaluation

Overview

A clone of the upstream open-source caesarHQ textSQL repository, retained for evaluation. It contains the original maintainers’ full history (~1,200 commits) with no studio-specific modifications, used as a reference and benchmark alongside the adapted textSQL fork.

Why It Exists

To study the upstream text-to-SQL implementation in depth, its prompting strategy, few-shot table selection, and failure-recovery logic, before and during the studio’s own adaptation work.

What We Built

Nothing custom here: this is the unmodified upstream project (Flask API + React client) kept as the canonical reference. The interesting parts are the upstream techniques observed, few-shot table selection, comment stripping during table selection, prompts tuned to avoid forcing CTEs, and “think-through on failure” retry behavior.

Technologies & Approach

Python/Flask + React with OpenAI GPT models for NL→SQL. Reviewed for its prompt engineering and reliability patterns, which informed the studio’s parallel adaptation.

Outcome / Impact

Served as a reference implementation and evaluation baseline. Captured here for completeness and to document the studio’s diligence in studying production-grade OSS before adapting it.

Capabilities Demonstrated

• Evaluating and learning from production OSS LLM tooling
• Understanding NL-to-SQL prompt engineering and reliability patterns