Not every search calls Jev
The source guards paid-plan/feature eligibility and skips small pools or clear winners. Comparing only reranked examples would omit the larger routing policy.
Inspect the supporting source ↗Optional second-stage search: widen the hybrid pool, then Score-rerank candidates with Workers AI typesafe/jev.
Sends a Score question per candidate, reorders and drops low scores; the model id is typesafe/jev.
Attaches Jev rerank to existing MCP search instead of replacing retrieval. License is Fair Source FSL-1.1-ALv2, not OSI open source.
Pinned public source was reviewed. The license is FSL-1.1-ALv2. This site did not execute the project or independently measure performance or safety.
Add an optional ambiguity-sensitive second stage to Kody’s hybrid MCP search.
Material analysed:
The source guards paid-plan/feature eligibility and skips small pools or clear winners. Comparing only reranked examples would omit the larger routing policy.
Inspect the supporting source ↗The source defines a 0.45 mean-confidence floor and a 1.5 keep-score threshold on a 0–3 rubric. These are Kody policy choices, not universal Jev thresholds.
Inspect the supporting source ↗Useful when hybrid results are ambiguous and a compact candidate card contains enough evidence. Compare with memsearch when complete passage text matters more than small-card cost.
Read evaluateJevSearchNecessity and the batch constants in search-jev-rerank.ts. Follow a clear winner to the skip branch, then a malformed/incomplete batch to fallback rather than treating it as a successful rerank.