- SignalDesk5 hr ago
Original Summary
jurl is a CLI. Tell it what you want from a page and it picks exactly that out: the paragraphs that answer a question, the code, the links worth following, or a photo you describe. On Wikipedia's Cologne page it finds the cathedral among 73 images in 2.3 s; on the Jevons page, the two paragraphs out of ~240 that explain his paradox.<p>The link above lets you try it in the browser, no key needed. Or in your terminal, with your own TypeSafe key (it asks for it on first run):<p><pre><code> brew install rayoplateado/tap/jurl jurl -q "What is the Jevons paradox?" en.wikipedia.org/wiki/William_Stanley_Jevons jurl --code github.com/BurntSushi/ripgrep jurl -f "a cathedral" en.wikipedia.org/wiki/Cologne jurl -l -n 3 news.ycombinator.com | xargs -n1 jurl -n 1 </code></pre> How it works: jurl splits the page into blocks, links and images, and turns each one into a yes/no question ("does this block answer the question?"). A decision model, Jev from TypeSafe, scores them all in a single request. For images, Clef on Cloudflare Workers AI looks at the pixels. The models return probabilities, never text: jurl picks, it doesn't write. Everything it prints is on the page, so it's safe to pipe into xargs curl.<p>Most pages take under a second and cost a fraction of a cent, about a thousand pages per dollar, paid straight to the APIs: jurl calls them directly with your own keys (TypeSafe for text, plus Cloudflare only if you want it to look at images).<p>Written in Rust, MIT or Apache-2.0. Code: <a href="https://github.com/rayoplateado/jurl" rel="nofollow">https://github.com/rayoplateado/jurl</a><p>I'd especially like to hear about pages where it picks the wrong thing.<p>P.S. It's pronounced /xaɾl/, "harl": the name is a nod to the Spanish comedian Chiquito de la Calzada and his "jarl".
- 情报分类:技术学习与提效
- 分类依据:内容涉及技术、AI、软件工具或工程实践
- 信息来源:Hacker News 新项目
- 发布时间:2026/10/5 22:38:02
- No replies yet