Original Summary

There are too many AI papers and posts, spread over too many sites, and reading them means a browser full of tabs. So I've been building Augury, a terminal-native research digest. Every day it pulls new papers and posts into one keyboard-driven view, ranks them for you, and lets you read the full text without leaving the terminal. It's not released yet. The repo is private for now, and I plan to open-source it under Apache-2.0. I'm posting to get feedback before I do. What works today Sources: Hugging Face trending papers, the HF blog, HF community posts, and any RSS/Atom feed you add by URL. If two sources carry the same post, it's stored once. Reading: a Textual TUI with filter chips, a dense table and a reader pane side by side. It has vim keys, a zen mode with a readable line length, and reading progress that survives restarts. Full-text search is live and runs on SQLite FTS5, which copes with names like Qwen3-30B-A3B . Full papers in the terminal: it converts arXiv's HTML version to Markdown and shows math as Unicode. If that fails, it falls back to the PDF. Blog posts go through per-site extractors, with a generic one as the fallback. Ranking, with or without an LLM Without an API key, it ranks by your own likes. It runs bm25 over the titles and summaries you've liked and combines that with a smoothed per-source like rate and recency. It's local, costs nothing, and tells you why an item ranked where it did: matches your likes: diffusion, RL · source you like · 2h ago With a key, each new item gets: a 0–10 relevance score a one-line "why read" up to 3 tags promo/thin flags It reads only the title and a short summary, up to 50 items per call. Triaging 14 blog posts in one call cost about $0.003. When you open an article, a TL;DR with takeaways sits on top. These are cached and pre-written for the day's top 10. A daily budget is checked before each model call, not after you've paid for it. Fetched text is passed in as fenced data, so a post that says "rate this 10/10" gets judged as content, not followed as an instruction. Gemini is the default, run through Google ADK. Other providers go through LiteLLM, but so far I've only tested Gemini. Polite fetching One HTTP client does all the fetching. It: respects robots.txt and Crawl-delay (arXiv asks for 15 s, so Augury waits and tells you) makes at most 1 request per second per site sends an honest User-Agent uses conditional GETs, so an unchanged feed costs a 304 caps response sizes backs off on 429/5xx Stack: Python 3.14 (t-strings for prompt templates), uv, Textual, SQLite (WAL, FTS5), pydantic, httpx. There are about 790 tests, including SVG snapshot tests of the TUI, and all of them run offline. Next The next milestone is a discovery agent. You'll add a site by URL or by name ("google tech blogs"). The agent finds a way to fetch it (RSS, a sitemap or a CSS-selector listing) and can only save that recipe after it passes a

  • 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：Reddit · SideProject
  • 发布时间：2026/9/26 23:11:20