Jev Reranker (Rust CLI)
JSON-in/JSON-out CLI that reranks, filters or compresses search results with independent Jev Noul checks, keeping metadata intact.
Overview
Jev Reranker (shinpr) is a command-line tool that chooses which search results are worth passing to an LLM. It reads a JSON array of candidates from stdin and writes a JSON array to stdout, so it fits after BM25, vector search or any command that emits JSON. You choose the text field and an optional context field such as a title; IDs, source paths and retrieval scores pass through unchanged, and existing retrieval scores do not affect Jev's judgment.
Three modes are available: rerank puts relevant candidates first with a 0-1 rerankScore; filter removes candidates with no usable evidence and keeps input order with an evidenceScore; compress extracts query-specific passages to send shorter text to the LLM. Each decision is a separate Noul check, and thresholds and ordering are applied in code. The README notes that reranking BM25's top 100 rather than top 30 on FiQA raised nDCG@10 from 0.36 to 0.40, at one extra request per 30 candidates. It is distributed through npm for macOS, Linux and Windows.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this candidate relevant to the query? | yes · no |
| Yes / no | Does this candidate contain usable evidence? | yes · no |
| Yes / no | Should this piece of source text be retained for the query? | yes · no |
How it works
- Pipe a JSON array of candidates into the CLI
- Choose rerank, filter or compress mode
- Jev answers a Noul per candidate (or text span)
- Output the same objects with scores, ordered or trimmed
Strengths
- Composable with any search command
- Metadata preserved untouched
- Three modes for different context needs
Caveats
- Each 30 candidates adds an API request
- Requires a TypeSafe key
- Unofficial tool
Getting started
npm install --global jev-reranker, export TYPESAFE_API_KEY, then pipe JSON into jev-reranker --query "..."At a glance
- Type
- CLI
- Status
- beta
- Language
- Rust
- Runtime
- CLI distributed via npm
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Retrieval and RAG: JSON-in/JSON-out CLI that uses separate Jev `Noul` checks to rank candidates, apply evidence thresholds, or extract source text while keeping those decisions independent.
yibie/awesome-jevUnofficial JSON-in/JSON-out CLI that uses Jev `Noul` judgments to rerank search results, filter documents without usable evidence, or extract query-specific passages
AnotiaWang/awesome-decision-modelsRerank, filter, and compress JSON search results with TypeSafe AI's Jev.
logicrw/awesome-jev-projectsRust JSON-in/JSON-out CLI that asks Jev about relevance, usable evidence, or which source text to retain, then applies the resulting order and thresholds in code.
cobanov/awesome-jevRerank, filter, and compress JSON search results with TypeSafe AI's Jev.
hellogumbo/awesome-jev
Frequently asked questions
What modes are there?
rerank, filter (remove candidates without usable evidence) and compress (extract query-specific passages).
Does it change my metadata?
No. Other fields pass through unchanged.
How many candidates should I pass?
More than you keep; the README shows gains from reranking a top-100 shortlist.
Related projects
Long-document search that walks a PageIndex tree, asking Jev Choice which page or branch answers the question, with no embeddings.
ProjectprototypePythonBrowser PDF search by meaning: Jev answers a yes/no per line on whether it answers your query, and matching lines light up.
ProxyprototypeTypeScriptPostgreSQL toolkit with 41 Jev-powered semantic operators and natural-language-to-SQL, leaving joins and arithmetic to SQL.
ProjectbetaPythonTypeSafe cookbook showing how to evaluate many typed questions over one shared state in a single Jev request.
Articleproduction- Jev Search★ 534
Plain-language web search where a decision model like Jev picks query, sources and time range, then ranks Search1API results.
AppproductionTypeScript MCP server over Apple developer docs and WWDC transcripts that uses hybrid search, then Jev relevance scores to rank results.
MCP serverproductionTypeScript