QueryLens launches with a useful benchmark result: none yet

QueryLens arrived as a work-in-progress text-to-SQL application with an unusually clear line between what its repository implements and what it has not yet proved. The project’s README says that every benchmark number will come from evaluation output files and, more importantly, that none of those numbers has been measured yet.

That disclaimer is the most useful part of the launch. New analytics-agent repositories often put benchmark names beside architectural claims before readers can inspect a completed run. QueryLens instead describes a planned BIRD mini-dev evaluation with ablations for schema retrieval and self-correction, while withholding an accuracy claim until outputs exist. It is not evidence that the system works well; it is evidence that the author knows the difference between an evaluation plan and a result.

Safety is split into layers

The implementation is aimed at PostgreSQL and pairs a FastAPI backend with a Vue 3 interface. Its documented execution boundary combines a read-only database role, read-only transactions, timeouts and an AST validator. The README explicitly says the prompt is not a security layer. That matters because a model instruction such as “only write SELECT statements” is not an enforcement mechanism: QueryLens’s design puts several controls between generated SQL and the database.

The repository also describes local-embedding schema retrieval through pgvector, self-correction, Redis caching and rate limiting. Its interface is intended to expose the generated SQL alongside a result table and chart, rather than returning only a prose answer. Follow-up questions are part of the planned interaction model.

What is actually available

At inspection on September 30, the public repository showed 22 commits, no tagged releases, no stars and no forks. GitHub displayed no license for the project. That means the code is inspectable, but users should not infer permission to reuse or redistribute it until the maintainer adds explicit terms.

The near-term test is straightforward: publish the promised BIRD output files, preserve model replies so the run can be audited, and show the ablations separately. Until then, QueryLens belongs in the directory as an inspectable new entrant—not as a benchmarked one.

That distinction is worth preserving. A blank scorecard with a reproducible plan is more informative than a polished percentage without artifacts.