Using semantic benchmarks to build a self-improving text-to-query agent | Dark Hacker News