Hev Reranker

Partly. Paired bootstrap per query puts Jev at or above Cohere rerank-v3.5 on all three corpora, two wins and one tie, and above Mixedbread with one interval touching zero. Against Voyage rerank-3, the strongest purpose-built reranker in the table, it trades: an edge on SciFact, a tie on NFCorpus, and a loss on FiQA by about 0.02 nDCG@10. Voyage is as good and slightly cheaper.

Yes. The limits section says hosted only, and that the data leaves your environment. Nothing in the README documents an on-premises, self-hosted, or OpenRouter endpoint for this recipe, so if your documents cannot go to a third-party API this does not fit as written, even though the prompt, state schema and wrapper are all included.

Both. A request caps at roughly 32k tokens for state plus questions, about 30 to 50 typical passages per call, and longer lists are split into concurrent calls. Rate limits are undocumented and sustained 429s appeared at about 24 requests in flight. Tail latency on 30-document states is p95 0.8-1.8 s, higher than specialized hosted rerankers; per-pair calls trade 30 times the calls for p95 near 0.25 s.

Yes, and that is the calibration payoff of scoring with a probability rather than a relevance label. On the SciFact run, documents scored at or above 0.9 were judged relevant 76% of the time, and documents scored below 0.1 were relevant 0.5% of the time. Passing threshold=0.1 drops candidates the model is confident are irrelevant, and scores stay comparable across split calls because each is an absolute probability.

pip install hev-rerank or uv add hev-rerank, with TYPESAFE_API_KEY in the environment. The rerank function takes a query string and a list of documents, and returns RerankResult(index, score, document) best-first. top_n trims the list and threshold drops candidates the model is confident are irrelevant, so the same call both reranks and prunes.

About $0.13 to $0.19 per 300 queries for the 30-Nouls-in-one-state shape against roughly $0.60 for Cohere rerank-v3.5, $1.05 for Mixedbread mxbai-rerank-large-v2, and about $0.15 for Voyage rerank-3. Claude Opus 5 as a listwise judge runs to $28.40 over the same slice. Those are the runs' actual token usage at list price, not list prices alone.

Because a reranker that does not know the corpus should not assume how many documents are relevant. A single Choice is slightly better when exactly one document is relevant, worth about +0.01 nDCG@10 on SciFact, and much worse when many are: NFCorpus has a median 16 relevant per query and scores 0.316 against 0.358 for per-document Nouls. All 30 questions in one request are evaluated in parallel, so 30 documents cost roughly the wall time of 5.