Awesome Search KG

hev-rerank

2 min read

hev-rerank

hev-rerank is a small open-source Python wrapper that uses TypeSafe’s Jev model as a search reranker: one call carries the query and up to 30 candidate documents, and a true-or-false question (Jev’s Noul type) is answered per document as a calibrated probability. The ranking is the sort by that probability.

It was released by Hev alongside the benchmark in Hev meets Jev, deliberately as a minimal reference implementation rather than a framework — the stated contents are the prompt, the request schema, a 90-line wrapper handling chunking and a prune threshold, and the evaluation results with confidence intervals.

Repository: https://github.com/hev/reranker · PyPI: https://pypi.org/project/hev-rerank/

Usage

# pip install hev-rerank
from hev_rerank import rerank
 
hits = rerank(query, shortlist, top_n=10, threshold=0.1)

The threshold argument is the part that distinguishes this from a conventional reranker client: because the score is a probability rather than a Cross-Encoder logit, an absolute cutoff is portable across corpora, and a single call performs a rerank and a prune in the same round trip.

Positioning

The author’s own read, published with the code, is not a win claim. Voyage rerank-3 scored marginally higher and cost marginally less in the same benchmark, and the batch call shape has a worse latency tail (p95 near 1.4 s for thirty documents in one call). What the approach adds is the calibrated probability and the fact that the same call shape also covers routing, classification, and gating — and that it requires no reranker training at all.

  • Jev — the model this wraps
  • TypeSafe — the model’s vendor
  • Hev meets Jev — the benchmark and the release announcement
  • Hev — the author
  • BEIR — the evaluation suite

Graph View

Backlinks