Same Material, Better Order: Reranking Lifts Correct-Passage Hit Rate to 90%

mikegiannulis · x · 2026-09-18

The thread reveals the test methodology: pulling real inputs from their own database, comparing against past Claude decisions and human labels or deterministic ground truth, and setting pass/fail rules before seeing results — a pure offline replay, no live customer experiment.

Headline result: for book draft source retrieval, embedding similarity ranked the correct passage first 60% of the time; Jev reranking hit 90%. Same material, better order — a much better starting point for writing.

Related event: Small model beats LLM on reranking and routing; real win is latency, not the 20-cent bill(8 posts)→

Original post →

More from coding & agent

coding & agent channel →