Repository navigation
RepoFind v0.2.0
Semantic search now works without any API key: RepoFind ships with a local embedding model that runs inside the app. There is also a new Atlas page that sorts your repos by topic, language and recency.
What's new
- Local embedding models. The default,
multilingual-e5-small(135 MB), downloads the first time you build the index. After that, search works offline and nothing leaves your machine. You can also downloadbge-small-zh-v1.5,multilingual-e5-baseorbge-m3in Settings, switch between them, or delete them. Downloads resume after an interruption and fall back to hf-mirror.com when huggingface.co can't be reached. - Three embedding sources. Pick a local model, a server already running on your machine (LM Studio, llama.cpp, Ollama), or any OpenAI-compatible API. If you configured an API in v0.1.x, the app keeps using it.
- Atlas, now the first page. It counts your repos by topic, language, last touched, local commits, source host and owner. Click any count to filter; search then runs only within the filtered repos. A radar view puts each repo in a sector by topic, farther from the centre the longer since you last touched it.
- Open in Finder / Terminal. Every result can be opened in Finder (File Explorer or the file manager on other systems) or in a terminal. On macOS you can choose Terminal, iTerm, Ghostty, Warp and others in Settings.
- Faster. Rebuilding the index only re-embeds repos whose text changed. Topics are worked out once, when you scan. On 328 repos, opening the Atlas went from about 740 ms to under 30 ms, and result lists are virtualized.
- Richer index. It now reads each repo's description, keywords and dependencies from its manifest (package.json, Cargo.toml, pyproject.toml, go.mod, pom.xml and others), its top-level files, and how many commits you made locally.
- Clearer empty states. If search finds nothing because there is no index yet, the app says so and offers Build index now.
Downloads
| Platform | File | Needs | Tested |
|---|---|---|---|
| macOS, Apple silicon | RepoFind-0.2.0-macos-arm64.zip |
macOS 14+ | CI, plus a local build of the same code run by hand |
| macOS, Intel | RepoFind-0.2.0-macos-x86_64.zip |
macOS 14+ | CI only |
| Linux x86_64 | RepoFind-0.2.0-linux-x86_64.tar.gz |
glibc 2.39+, GTK 4.14+, WebKitGTK 6.0 (Ubuntu 24.04 or later) | CI only |
| Linux arm64 | RepoFind-0.2.0-linux-arm64.tar.gz |
same as above | CI only |
| Windows x86_64 | RepoFind-0.2.0-windows-x86_64.zip |
Windows 10 1809+ | CI only |
"CI only" means the type check and headless tests pass and the package builds on that platform; nobody has opened the app there yet. On Apple silicon, downloading the default model, indexing and semantic search were run by hand on real repos. The Open in Finder / Open in Terminal buttons are covered by tests but have not been clicked on any real machine yet.
The Intel Mac build has no local models, because ONNX Runtime ships no prebuilt library for it. Use a local service or an external API there; Settings says so.
Each file has a .sha256 next to it.
Upgrading from v0.1.x
- Your scan roots and API settings carry over.
- Rebuild the index once. Older indexes lack descriptions and topics, and older vectors don't record what text they were made from, so the first rebuild embeds everything again.
- v0.1.x can't read vectors written by v0.2.0. If you go back, that version falls back to lexical search until you rebuild with it.
Install
- macOS: unzip, then move
RepoFind.appto Applications. The build is ad hoc signed and not notarized, so clear the quarantine flag once:xattr -dr com.apple.quarantine /Applications/RepoFind.app - Linux: extract the archive and run the
RepoFindexecutable inside it. - Windows: unzip and run
RepoFind.exe. It is not code signed, so SmartScreen warns on first launch ("More info" → "Run anyway").
现在不配 API key 也能语义搜索了:RepoFind 内置了在应用里运行的本地 embedding 模型。另外新增了「总览」页,按主题、语言和最近变动给仓库分类。
新功能
- 本地 embedding 模型:默认模型
multilingual-e5-small(135 MB)在第一次建索引时自动下载,之后离线也能搜,数据不出本机。设置里还可以下载bge-small-zh-v1.5、multilingual-e5-base、bge-m3,随时切换或删除。下载中断能续传,连不上 huggingface.co 时会自动改用 hf-mirror.com。 - 三种向量来源:可以选本地模型、本机已经在跑的服务(LM Studio、llama.cpp、Ollama),或任意 OpenAI 兼容接口。v0.1.x 里配过 API 的,升级后继续沿用。
- 总览页,现在是首页:按主题、语言、最近变动、本机提交、来源、归属统计仓库。点任意一项就筛选,之后的搜索只在筛出来的仓库里进行。雷达图里每个仓库是一个点:所在扇区是主题,离圆心越远说明越久没动过。
- 用 Finder 或终端打开:每条结果都能在 Finder(其他系统上是资源管理器或文件管理器)里打开,也能在终端里打开。macOS 上可以在设置里选 Terminal、iTerm、Ghostty、Warp 等。
- 更快:重建索引时,只有文本变了的仓库才重新生成向量;主题在扫描时算一次。在 328 个仓库上,打开总览从约 740 ms 降到 30 ms 以内,结果列表也改成了虚拟列表。
- 索引内容更丰富:会读 manifest(package.json、Cargo.toml、pyproject.toml、go.mod、pom.xml 等)里的简介、关键词和依赖,以及仓库根目录的文件,并统计本机提交次数。
- 空状态说得更清楚:还没建索引就搜索时,应用会直接说明原因,并给出「立即建索引」按钮。
测试情况
表里标 "CI only" 的平台,只保证类型检查、无头测试和打包在该平台通过,还没有人实际打开过。Apple 芯片上用真实仓库手动跑过:下载默认模型、建索引、语义搜索。「在 Finder / 终端中打开」两个按钮有测试覆盖,但还没在任何真机上实际点过。
Intel Mac 版不带本地模型,因为 ONNX Runtime 没有提供它的预编译库。在 Intel Mac 上请改用本地服务或外部 API,设置页里会有提示。
从 v0.1.x 升级
- 扫描目录和 API 配置都会保留。
- 升级后需要重建一次索引:旧索引缺简介和主题,旧向量也没记录生成时用的文本,所以第一次重建会重新生成全部向量。
- v0.1.x 读不了 v0.2.0 写的向量。如果退回旧版,搜索会退回词法检索,用旧版重建一次索引才能恢复。
安装
macOS 解压后拖进「应用程序」,并执行一次 xattr -dr com.apple.quarantine /Applications/RepoFind.app;Linux 解压后运行目录里的 RepoFind(需要 glibc 2.39+、GTK 4.14+、WebKitGTK 6.0);Windows 解压后运行 RepoFind.exe(没有签名,SmartScreen 会提示,点「更多信息 → 仍要运行」)。