Jev 101
Start Guides
Search
中文 EN
Retrieval heat 2 · Cool★ 0heat incomplete (no X)

Jev-Reranker

A head-to-head test of Jev and BGE as RAG re-rankers on 300 multi-hop HotpotQA questions, with the same search, generator and questions for every variant.

Repo: subhan75/Jev-Reranker ↗

What it is

A small, honest benchmark of Jev as a re-ranker.

What it does: Compares no re-ranking, BGE and Jev on the same RAG pipeline.

Why it belongs in the showcase: Reports cost, latency and confidence intervals, not just a headline.

Caveats: Single dataset.

Evidence

Public MIT repo, created 2026-10-21. Jev (typesafe/jev-1.14 via OpenRouter) asks one Noul per passage and ranks the top 30 by probability. Both needed sources in the top 5: 69% with no re-ranking, 86% with BGE, 90% with Jev. Answer quality score 60, 68 and 73. Re-rank time 0.24 s per question for Jev against 8 s for BGE on a CPU laptop; Jev costs about $0.25 per 1,000 questions.

Caveats

  • One dataset, one run of 300 questions
  • Only the 4-point answer-quality lead is statistically solid, per the author
  • BGE was timed on a CPU; a GPU would be much faster

Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.