hev reranker
Uses Jev as a reranker: one relevance question per document over a BM25 shortlist.
Uses Jev as a reranker: one relevance question per document over a BM25 shortlist.
One state carries the query and up to 30 candidate documents keyed D00, D01 and so on. Each document gets its own noul question in the same request, and the calibrated probability works as both the sort key and an absolute pruning threshold.
Does this document contain information that answers or directly addresses the query?
The state is a query plus documents keyed D00, D01 and so on from a BM25 shortlist. The repository defines one domain-neutral noul question per document, with true and false criteria; probabilities above the threshold are kept.
A lexical retriever such as BM25 returns the top candidates for the query.
The query and the shortlisted documents are packed into one state keyed D00, D01 and so on.
Every document receives the same relevance question, answered in parallel inside one request.
The probabilities reorder the shortlist and a threshold removes weak candidates before they reach the next stage.
A reranker usually needs a trained scoring model; here one prompt and a schema turn a decision model into one, with a threshold that has meaning rather than a relative score. The repository reports nDCG@10 across three BEIR collections against hosted rerankers and LLM judges.