jev rerank

Score every candidate's relevance to a query independently, then sort. Several can be relevant, or none.

jev rerank <query> (--files <paths...> | --candidates <ref> | --lines <ref>) [options]

When to use

In front of any search: take the top 30 from BM25 or a vector index, keep what Jev says is relevant. The cookbook shows top-1 accuracy rising from 5% to 18% with one judgment per pair. Where find asks "which one is best", rerank asks "is this one relevant" for each.

Options

Option	Meaning	Default
<query>	What results should be relevant to	required
-f/-c/-l	Candidates, same forms as find
-k, --top-k <n>	How many ranked results to return	10
--min <p>	Relevance at or above which a candidate is kept	0.5
--criteria <text>	What counts as relevant, e.g. "a passage stating the rule, not commentary"
--fail-on <list>	Exit 2 on empty (nothing kept)	none

Output

ranked[] with id, relevance, kept, text; kept[] ids. Up to 250 candidates per call.

Example

my-search "$q" --json | jev rerank "$q" -c - --min 0.6 --json | jq '.kept'

2 of 5 shown are relevant (min 0.6)  statute of limitations for contract claims

#  Rel        Id    Text
-  ----  ----  ----  ------------------------------------------------------------
1  0.94  keep  p17   An action for breach of a written contract must be brought…
2  0.71  keep  p03   The limitations period begins when the breach occurs…
3  0.22  drop  p09   Courts have discussed equitable tolling in other contexts…

Pattern: reranking cookbook.