Skip to content

Jev Retrieval

A Rust CLI that finds files and line ranges with two ingredients: a stateless in-memory BM25 pass for recall and Jev calibrated judgments for precision. No embeddings, no vector database, no index directory, no repository state.

Stars
6
Language
Rust
License
APACHE-2.0
Published by
Community
Author
romeromarcelo
Last verified
Oct 2026

About this project

Find files and content using Jev's calibrated decisions. 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.

Quoted from the project's README. Read the full README.

Highlights

  • Grep-shaped: positional query plus path, path:line output with a 0..1 score, and grep exit codes.
  • No persistent state between runs - BM25 is rebuilt in memory each invocation, about 1s at repo scale.
  • Responses cache under ~/.cache/jevr keyed on request bytes including file content, so entries cannot go stale.
  • Reports 0.900 any-gold@10 on SWE-bench Lite file localization against BM25@27K at 0.513.

Watch out

Requires Rust 1.88+ and a TypeSafe key at runtime, so every cache-cold query is a billed call - the README puts it at roughly 2 seconds and about a cent per cold query, with a warm re-query measured at about 0.1s and zero billed requests. It is not a reranker frontier model everywhere: on BEIR NFCorpus it scores 0.363 nDCG@10 against the best zero-shot reranker's 0.399, and the README says full-corpus retrieval is bounded by BM25 candidate recall on vocabulary-divergent domains, which is why a candidate_cap: 100 overlay is the measured recommendation. Hand-tuned -k keyword variants measured worse than the plain question (macro-F1 0.614 vs 0.673), so the obvious tuning knob is a downgrade. Results are a listwise rerank when that succeeds and score order otherwise, so ordering behaviour changes with the mode. This README copy is cut off partway through the flags table, so the remaining flags and sections after it were not visible here. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

72GitHub stars
MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.
Jev Agents & MCP#community#jevbest#catalog-import
35GitHub stars
Rust terminal tool that asks Jev noul, choice and score questions about piped text or files and returns calibrated probabilities with stable exit codes for shell scripts, CI jobs and agents.
Jev Agents & MCP#community#jevbest#catalog-import
35GitHub stars
Diff-aware test selector that sends the git diff to Jev with one question per test, then prints the native filter flags and hands them to vitest, jest, node, bun, Playwright, cargo, or go test.
Jev Agents & MCP#community#jevbest#catalog-import
22GitHub stars
Sentence-driven runner where one command is matched to installed reflex programs, planned up front with per-step probabilities and gates.
Jev Agents & MCP#community#jevbest#catalog-import
9GitHub stars
Hermes Agent plugin exposing Jev as four coding-workflow tools for parallel judgments, binary checks, option routing and rubric scoring, with a bundled question-design skill.
Jev Agents & MCP#community#jevbest#catalog-import
LangChain deepagents middleware that replaces SkillsMiddleware with per-turn judging, loading only the SKILL.md files a turn needs so large catalogs stay out of the prompt.
Jev Agents & MCP#community#jevbest#catalog-import
Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai