📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
hn · Show HN · 39 分 ← 返回列表

Show HN: Local embedding and ranking for tech feeds

I follow a bunch of blogs and feeds, and Skimfeed was too busy for me. Initially I built a Glance page, and when colleagues saw it they asked me to share it. That led to (the idea of) okaeri, a minimal startpage & feed recommendation algorithm that runs locally. In the background it runs EmbeddingGemma 2 (270M, 8-bit ONNX, on CPU) to embed new stories from the feeds you follow, and a custom ranking algorithm uses those embeddings (plus what you open, like and hide) to put the most relevant items

五维评分

付费意愿
28
需求真实度
35
个人可落地
72
竞争空间
32
新鲜度
55

AI 分析

痛点:技术博客和RSS订阅源太多太吵,用户想要一个本地运行、可查看可手动调节的个性化排序推荐,而不是把阅读行为数据交给云端算法。

目标用户:隐私敏感的技术信息重度订阅者,即开发者、研究者和独立黑客等RSS与博客重度用户。

证据:when colleagues saw it they asked me to share it;What I like about this is having a recommendation algorithm that runs on my own machine, and I can see and adjust what it does.;Some future ideas are: Creating a self hosted version you can easily sync to your phone / Building & hosting a public version.

变现:开源核心保持AGPL免费,对托管版和手机同步、跨设备向量同步收取订阅费;或提供一次性买断的桌面与原生客户端、企业内网部署授权,并辅以GitHub Sponsors或打赏。

MVP:两周内做一个本地优先的极简版:使用一个现成的RSS解析器和更小的嵌入模型(或采用本地Ollama/API混合方案),用SQLite存256维向量;只提供“按点赞、点击、隐藏做加权排序”的排序逻辑,并提供3个可调滑杆;界面采用PWA或单页startpage;不做账号系统和云同步。

竞品:Feedly Leo AI、Inoreader、Readwise Reader、Skimfeed、NewsBlur、Reeder

摘要:作者因嫌Skimfeed太杂,自建了okaeri:本地用EmbeddingGemma 2(270M、8bit ONNX、CPU推理)为订阅源做嵌入,向量截断到256维存SQLite,嵌入时峰值内存约1.3GB,按主题分别记录偏好并用滑杆调节排序,AGPL开源,支持macOS、Linux和Windows。同事看到后主动要求分享,说明有一定真实需求,但Hacker News仅5 points、0评论,无任何付费或收入证据,作者自己的下一步想法是手机同步和托管公开版。