Jev Reranker
Filter and rerank retrieval candidates using typed Jev relevance judgments.
Added to Jevfast
How it uses Jev
The reranker sends a query and candidate documents to Jev for Noul judgments. In pointwise mode, each request contains one document with the query and a `relevant` question asking whether that document helps answer the query. The library also supports pairwise and listwise payloads. It validates each requested Noul answer, then returns above-threshold documents in descending score order, capped by `top_k` when set. Listwise ranking splits candidates when estimated request or state budgets are exceeded.
What Jev decides
Pointwise document relevance (illustrative request)Example answers · not a recorded Jev response · Source ↗
Question 1 · relevant
YOUR APP
INSTRUCTION
Does this document help answer the query?
STATE
The request contains the caller's query and one candidate document as `query` and `document`. No example query or document text is included in the saved source.
JEV · NOUL
YesNo
App workflow
- Prepare a query and candidates
Pass the query and candidate documents to the selected pointwise, pairwise, or listwise mode.
- Build and send Jev questions
Pointwise mode sends one document with a `relevant` Noul question; the client validates one answer for each requested key.
- Split requests that exceed budgets
Listwise mode estimates request and state sizes, partitions candidate indices when needed, and retries splitting after a provider context-limit response.
- Return ranked documents
Keep scores at or above the threshold, sort them from highest to lowest, and cap the returned list at `top_k` when configured.
Why it is interesting
The reranker treats relevance as evidence for answering a query and supports independent document scoring as well as pairwise comparison. Its listwise partitioner responds to context limits by splitting the candidate set, which preserves a larger ranking job when the full request cannot fit.