Jev Reranking

An independent evaluation of Jev in three retrieval stacks, with recalculated recall, latency, and cost against DeepSeek, GPT-4o-mini, and GPT-5-mini baselines.

Reranking benchmark that gave Jev, Cohere Rerank 4 Pro, ZeroEntropy zerank-2, and DeepSeek the same thirty BM25 candidates across eight English datasets, publishing saved responses, scoring code, and paired-bootstrap intervals. Jev's rubric scored 0.692 nDCG@10 against Cohere Pro's 0.691.

Unofficial LlamaIndex reranker and router built on the official Python SDK: JevRerank scores retrieved passages and JevSingleSelector chooses which tool handles a query.

A CLI that reranks, filters, or extractively compresses JSON retrieval results by asking Jev separate Noul questions, then lets Rust apply the order and thresholds, reading stdin and writing stdout.

A graded relevance evaluation of Jev as a reranker: 9,831 labelled pairs from 164 queries over a 33,047-item skills catalog, comparing Jev score reranks with BM25, bge-m3, and rank fusion.

A sort CLI where the key is a plain-English description: Jev compares texts in pairs, jsort fits a Bradley-Terry scale to the answers, and prints each score with a standard error.

A Python library that reranks search results and filters retrieved documents with Jev, scoring each candidate's usefulness as evidence and dropping those below a configurable threshold before RAG context is assembled.

A local SQLite memory layer for CLI agents with an opt-in Jev reranker that batches Noul judgments over the top 40 recalled memories and falls back to a local cross-encoder on errors.

A cross-platform memory layer for AI coding agents: conversations from Claude Code, Codex, DeepSeek Harness, OpenClaw, and OpenCode become Markdown memories indexed in Milvus, with optional Jev reranking through the TypeSafe API.

A measured search reranking run over 33,047 catalog entries, 164 real queries, and 9,831 labelled pairs, reporting how Jev compares with BM25 and bge-m3.

Rank 182 agent skills in one request and re-read the top three in a second: over 488 requests, wrong skill loads fell from 16.8% to 7.3% and needless loads from 9.8% to 4.0%.

One question per query-candidate pair raises top-1 accuracy from 5% to 18% and top-10 accuracy from 38% to 62% on 40 CLERC legal queries.