Jev in production › Search, ranking and matching

jevr (jev-retrieval)

CLI pairs BM25 recall with Jev calibrated relevance scores to rerank search results; 0.900 any-gold@10 on SWE-bench Lite (vendor).

Open on GitHub ↗

0.900measured, as published by the source
Use
Search, ranking and matching
Industry
Developer tools
Form
Open-source tool
Stage
In production
Listed
2026-10-12
Found via
github
Repository
romeromarcelo/jev-retrieval
Stars
7
Forks
0
Last push
2026-10-12
Language
Rust
License
Apache-2.0
jevr (jev-retrieval) screenshot
docs/assets/demo.gif in the romeromarcelo/jev-retrieval README, Apache-2.0; shown from GitHub.

The README opens with

jevr is a simple, lightweight, blazing-fast information-retrieval CLI built from exactly two ingredients: a local, stateless BM25 pass for recall, and TypeSafe Jev calibrated judgments for precision. No embeddings, no vector database, no index directory, no repository state — point it at any codebase or plain-text document corpus (markdown, rst, txt, html, org, tex, ...) and ask a question.

Ask "where is the websocket reconnect logic?" or "how do I configure retry backoff?" and get back a ranked list of files and line ranges, each with a calibrated relevance probability — ready to open at the right spot. That recipe scores 0.900 any-gold@10 on SWE-bench Lite file localization, beats published BM25 and cross-encoder baselines on BEIR, and places...

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for search, ranking and matching