XERJ
Project / 项目 · In the catalogue / 在目录中查看
What Jev decides here · Jev 在这里做什么决定
noul: read in the code by a model, not yet by a person 由模型在代码里读到，尚无人复核
a model read the call site; no person yet 模型读过调用点，尚无人复核 · · Cited file / 引用的文件: engine/crates/xerj-rerank/src/lib.rs
Summary · 摘要
Local-first search engine and MCP server in one Rust binary. Its optional `rerank` stage asks Jev one `noul` per top hit (does `document` answer the query in the state?) and re-sorts by the probabilities. The authors report FiQA nDCG@10 rising from 0.24 to 0.36, about $0.64 for 648 searches.written by a model
一个本地优先的搜索引擎兼 MCP 服务器，打包为单个 Rust 二进制文件。可选的 `rerank` 阶段对每条靠前结果向 Jev 提一个 `noul`（`document` 是否回答了 state 中的查询？），再按概率重新排序。作者报告 FiQA 上的 nDCG@10 从 0.24 升至 0.36，总共 648 次搜索花费约 0.64 美元。模型撰写机翻
Notes · 备注
The queue cited a test double (scripts/bench/lib/jev_stub.py); this row cites the provider client. Reranking needs a TypeSafe key and sends the top hits' text to TypeSafe. XERJ's own /v1/systemone answer (xerj-history-vote-1) is not Jev. The authors found the probabilities uncalibrated (ECE 0.10–0.31). Row drafted from the discovery queue (issue #14) by an AI assistant that read the README and the cited file on 2026-10-26; a maintainer has not yet read the call site.
发现队列引用的是测试替身（scripts/bench/lib/jev_stub.py）；本条目改引提供商客户端。重排需要 TypeSafe 密钥，并会把靠前结果的文本发送给 TypeSafe。XERJ 自己对 /v1/systemone 的回答（xerj-history-vote-1）并不是 Jev。作者发现这些概率未经校准（ECE 0.10–0.31）。本条目由 AI 助手根据发现队列（issue #14）起草，于 2026-10-26 阅读了 README 和所引文件；维护者尚未亲自阅读调用点。
Launches · 发布记录
Not launched yet. 还没有发布过。 Launch it / 发布它 ↗ · How a product launches / 怎样发布一个产品
Makers · Maker
Made this? Confirm or correct its entry / 这是你做的？确认或更正条目 ↗
The same form takes the product's one line and a demo link, in its maker's own words. 同一个表单也可以补上产品的一句话和演示链接，用 Maker 自己的话。
Badge for its README · README 徽章
[](https://kydlikebtc.github.io/awesome-jev/products/xerj.html)One line for the product's README, linked to this page. The picture says that it is listed; while it is on the launch board, that it is launching; after a week in the top three, its place. It is rebuilt with the site, so the line never needs changing. 放进产品 README 的一行，链接到本页。图片显示已收录；在发布榜上时显示正在发布；进过某周前三后显示名次。它随网站一起重建，这一行不用再改。
Also under this pattern · 同一模式的其他产品
AutoGPT TypeSafe blocks
Seven production blocks — choice, score, yes/no, ask-many, route, pick-best, filter — with a UTF-8 byte budget, verbatim wire capture and eleven test files.
OpenViking: retrieval reranking
One Noul per candidate document in a single batched request, with the yes-probability used directly as the relevance score.
oh-my-pi
⌥ Coding agent with the IDE wired in
FastMCP jev_search transform
Two-stage MCP tool search: a wide Choice coarse-ranks the whole catalogue, then a shortlist gets full descriptions plus one Noul each to decide whether it does the job at all.
jcode: memory recall without embeddings
Replaces the whole retrieval stack for memory recall — no embeddings, no BM25, no reranker — with one batched Noul per candidate memory.