// the find
KKKKhazix/AIHOT
一个自己找热点、自己写日报的网站框架。把信源和精选标准换成你的，它就是你的行业热点站。
A self-hosted framework for running an industry news site, extracted from the AIHOT Chinese-language AI news site. It pulls from RSS, web lists, JSON feeds, X accounts, WeChat accounts and pushed content, scores items with an LLM, clusters coverage of the same event, and publishes daily, weekly and monthly digests. It is aimed at people who want their own Chinese-language industry site and are willing to replace the sources and selection criteria.
The selection prompts and thresholds sit in plain files (industry/prompts/ and industry/selection.ts), so changing the editorial standard does not mean changing code. That is the right split for a product whose value is the judgment. The heat score counts each independent source once per event over a 48-hour window and halves after 24 hours, so one outlet publishing ten articles does not push a story up the ranking. The calibration loop is concrete: you label your own items and run scripts/eval-selection.ts to see whether the scorer agrees with you, which most projects like this never provide. Output comes in several forms (RSS for full text and digests, a public API, an MCP server, llms.txt), and the deployment is Docker Compose with Postgres 17 and pg-boss, so the moving parts beyond the database are few.
This is an extraction from one production site, not a framework designed for reuse. The README says the code is synced from the live site, that there is no guarantee each sync stays compatible, and that the author is not a professional developer. Expect to read the code when something breaks. The pipeline makes several model calls per item (prefilter, two independent scoring passes, writing, translation, grouping with a review pass), and the top-level README gives no cost figure, so estimate the bill before pointing it at a few hundred sources. The bundled sources are overseas AI feeds, and the Chinese-language prompts and taxonomy are tuned to that domain. Moving to another industry means rewriting the taxonomy, topics, prompts and source list, which the customization guide frames as mostly content work but which still requires real editing and calibration. Scraping-based sources such as X and WeChat accounts are the most likely to need upkeep when the platforms change, and the README does not say how they hold up.