Small decisions. Interesting possibilities.Submit contentSubmit

How it uses Jev

When Jev reranking is enabled, Hippo sends the query and up to 40 numbered candidate memories, with candidate text redacted and truncated, and asks one typed `noul` question per candidate: the probability that it helps answer the query. Hippo validates every returned probability and sorts candidates by it; recall consumes that ranking downstream.

What Jev decides

Illustrative Jev decisionExample answers · not a recorded Jev response · Source ↗
Question 1 · c1
YOUR APP
INSTRUCTION

Does candidate 1 help answer the query? One Noul question is asked per candidate.

STATE

A query and numbered, redacted candidate memories from the project store, up to 40.

JEV · NOUL
YesNo

App workflow

  1. Retrieve and opt in

    `hippo recall "<query>" --reranker jev` selects the optional Jev reranker for the candidate pool; Jev uses the top 40 by default.

  2. Prepare candidate state

    Redact the query and candidate text, truncate each candidate to 1,200 characters, and number candidates in the state.

  3. Ask for relevance probabilities

    Send one `noul` question for each candidate, phrased as its probability of helping answer the query.

  4. Rank or fall back

    Require a finite 0–1 answer for every candidate, stable-sort by descending score, and use the local cross-encoder if the request or answers fail.

Why it is interesting

One batched call turns candidate memories into directly sortable relevance signals. The implementation also records that this helps ranking in measured evaluations but has not shown an answer-rate improvement over the local cross-encoder.