Decision: Jev reranks and gates those candidates.
I experimented with a Jev‑plus‑RAG variant where the retrieval stage stays identical to a standard RAG pipeline, but before the LLM sees any passages I invoke Jev to score each retrieved chunk for answerability. Each passage is asked a typed question like “Does this passage help answer the query?” and Jev returns a probability. I keep only those above a configurable threshold (e.g., 0.7) and also compute an aggregate answerability score for the gated set; if that falls below a second threshold I abort generation and return “not in the documents.” This gating layer replaces the usual relevance‑only reranker and directly targets hallucination reduction.
Decision: Use Jev for per‑candidate answerability scoring rather than a generic relevance reranker.
Decision: Apply a probability threshold to filter out weak candidates and a second aggregate threshold to optionally skip the LLM.
I captured several Jev‑style scoring use‑cases in my September 2026 X posts, each illustrating how deterministic candidate generation can be combined with semantic ranking.
Jevgrep code search Search a codebase by what the code does rather than its exact words. Jev scores candidate snippets and returns the most relevant code first.
Entity alignment Compare two candidate records and decide whether to merge, review, or keep them separate. Candidate generation remains deterministic while Jev handles semantic identity.
Retrieval reranking Let embeddings retrieve a broad candidate set, then use Jev to reorder passages by relevance. The generation model receives the most useful evidence first.
Memory promotion gate Capture a completed agent trace, then judge whether its corrections contain a reusable lesson. Trace-backed lessons can be promoted while task-specific noise is discarded.