AIHOT Open-Sourced: Build Your Own Industry Hot-Topic Site by Swapping Sources

What Is AIHOT?

AIHOT is a complete, production-grade website that turns a list of information sources into a curated hot-topic site. You provide sources (RSS, web lists, JSON APIs, X accounts, WeChat public accounts, or custom scripts), and the system automatically fetches content daily, pre-filters with an LLM, runs two independent scoring passes against the same criteria, clusters related reports into single events, ranks events by a heat algorithm that resists manipulation, and publishes a daily briefing at 8 AM plus weekly and monthly digests. The repository contains the full stack: frontend, backend, selection pipeline, clustering and heat algorithms, and — most importantly — every prompt and acceptance threshold in plain text.

Performance is notable: median page response is 10 ms, with 95% of requests under 50 ms, running on Node.js 24 and PostgreSQL without any external BaaS.

The Six-Step Selection Pipeline

Each piece of content passes through six stages:

Collection — pull raw items from configured sources.

Pre-filter — an LLM does a quick relevance pass.

First scoring — the same item is scored against the selection criteria.

Second scoring — an independent second pass with the identical prompt; both must clear the threshold.

Writing — accepted items get a Chinese title and summary.

Clustering & publication — vector search over the last two weeks finds candidate duplicates; an LLM judges whether they are the same event, a follow-up, or distinct. Follow-ups attach to the parent event, which then receives a synthesized overview. Heat is computed per event (not per article): each unique source counts once per 48 hours, with a 24-hour half-life, preventing volume spam from inflating rankings.

Three Key Design Decisions

1. Independent double scoring with the same rubric

LLM scoring is non-deterministic. AIHOT runs the identical prompt twice; only items that pass both passes enter the pool. Source-tier thresholds (official vs. self-media) are layered on top. Every prompt and threshold is exposed in the repo for line-by-line editing.

2. Event-level clustering

Multiple outlets covering the same story are merged. The system first uses title/summary embeddings to retrieve candidates from the prior two weeks, then asks an LLM to classify the relationship (same event, follow-up, or different). Uncertain cases are re-checked with a second model. The event page shows a consolidated narrative.

3. Heat by event, not by article

Each distinct source contributes at most once per 48-hour window; heat decays by 50% every 24 hours. A single publisher pushing ten articles or a bot farm reposting en masse cannot game the leaderboard. The author notes this anti-gaming rule is the "lifeline" of a site that decides daily what deserves attention.

Why Open-Source?

Professionals from law, HR, finance, precious metals, etc., asked the author (Kazik) to build industry-specific versions. His README answer: "I can't. I don't know your industry, which sources matter, or what counts as a hot topic. But you do. Since I can't serve everyone, I'm handing you the spark."

The core open-sourced asset is not the few thousand lines of TypeScript but the complete methodology: prompt texts, scoring rubrics, acceptance thresholds, and calibration procedures. The code is intentionally generic; all industry-specific logic lives under industry/, especially industry/prompts/selection-score.md.

Customization and Calibration

Adapting the site is designed to be trivial. The recommended workflow:

Feed the repo to an AI coding assistant (Claude Code, Codex) with a prompt like:

请读 AGENTS.md 和 docs/customize.md，把这个站改成「法律」行业的热点站。 我关心的是：……（你想盯哪些信源，什么消息重要、什么不重要，越具体越好）

Run the calibration loop: annotate a few hundred items, then execute scripts/eval-selection.ts to run SelectBench against your samples. Iterate the rubric until precision/recall meet your bar. Selection quality becomes measurable, not mystical.

Boundaries are clear: the repo ships with 18 overseas AI sources as demos, not the real AIHOT source list or operational data. The AIHOT name and logo are excluded from the MIT license — "swap in your own brand, and it becomes your site."

Deployment

One Docker-enabled machine plus an OpenAI-compatible API key (DeepSeek, Qwen, Zhipu, etc.) is enough:

git clone https://github.com/KKKKhazix/AIHOT.git myhot
cd myhot
node scripts/init-env.ts --llm-key <你的模型APIKey>
docker compose up -d --build

Visit http://localhost:3000; admin panel at /admin. Content starts appearing within minutes; the first full batch finishes in ~30 minutes.

Related Projects: Four Delivery Forms

The article positions AIHOT alongside three other open-source tools, forming a spectrum of delivery shapes for AI-processed information:

AIHOT (this project) — build a site for an audience .

Horizon (9,500+ stars) — push a personal bilingual briefing to WeChat, Lark, or email. Includes a built-in "AI creator" profile that breaks each story into event, timeliness hook, and angle suggestions.

last30days-skill-cn (1,800+ stars) — ask an agent "what has the AI circle discussed in the last 30 days?" It searches Weibo, Xiaohongshu, Bilibili, Zhihu, Douyin, etc. (8 platforms), producing a cited research report. Zero API keys except WeChat.

MediaCrawler (66,000 stars) — raw data layer for seven Chinese platforms using Playwright to handle login state and signatures; no JS reverse-engineering required. License permits learning/research only, no commercial use.

newspaper (15,000 stars) — the veteran extraction library (since 2013) that turns a URL into clean text, author, date, images, keywords. AIHOT, Horizon, and similar pipelines all rely on this step.

The choice depends on who consumes the output: a public audience, yourself, an agent, or your own custom pipeline.

Conclusion

Traditional open-source releases a library, framework, or component. Kazik has open-sourced an entire "how to run a media operation" playbook: source tiering, double-blind scoring, manipulation-resistant heat, and sample-driven calibration. The technical half is shipped; the domain-knowledge half is left for each industry's experts to fill in.