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。

What is AIHOT?

AIHOT is an open‑source AI‑driven news‑aggregation framework. It continuously pulls items from a configurable set of sources (RSS, web pages, JSON APIs, X/Twitter accounts, WeChat public accounts, or custom scripts), runs a large language model (LLM) over each item to filter, score, and rewrite it in Chinese, clusters articles that talk about the same event, and ranks those events by how many independent sources mention them. The result is a daily briefing (and weekly/monthly reports) that can be presented as a public website or consumed by other agents via RSS, an API, or plain‑text files.

Core workflow (six steps)

Ingestion – fetch raw items from the configured sources.
Deduplication & pre‑screen – drop exact duplicates and quickly discard irrelevant pieces.
Dual scoring – the LLM evaluates each item twice with the same scoring prompt; only items that pass both thresholds are kept.
Writing – the model generates a Chinese headline, a concise summary, a recommendation reason, tags, and (optionally) a translation of the full text.
Clustering – using vector similarity on the headline/summary, recent items are grouped; a second model decides whether they describe the same event or a follow‑up.
Hotness calculation – an event’s score is based on the number of distinct sources mentioning it (one count per source per 48 h, halved after 24 h). The top events become the daily “hot” list.

All prompts, thresholds, and selection logic live in the industry/prompts/ directory, so you can adapt the criteria without touching code.

What you get out of the box

Component	What it provides
Six source types	RSS, static URL list, JSON API, X/Twitter account, WeChat public account, custom script feed. Sources are tiered (official vs. media) and fetch frequency auto‑adjusts.
Selection pipeline	Pre‑filter, double LLM scoring, configurable thresholds, full prompt texts for transparency.
Automated writing	Chinese title, answer‑first summary, recommendation, tags, and optional full‑text translation.
Clustering & event view	Same story from multiple outlets merged into a single event page with a synthesized overview.
Hotness ranking	Event‑level popularity based on unique source count; fast‑rising events flagged as “↑”, brand‑new ones as “新”.
Reports	Daily briefing at 08:00, plus weekly and monthly reports, all split by categories with introductory text.
Search & topics	Company, direction, and content‑type taxonomy; full‑text and semantic search.
Agent‑friendly outputs	RSS feeds (selected, all, full‑text, daily), public REST API, `llms.txt` dump, and MCP format.
Admin backend	Source management, test crawls, diagnostics, per‑step model swapping, budget throttling, logs & alerts.
AI‑specific modules	Model leaderboard (aggregated public benchmarks) and Codex reset monitor; can be toggled per industry.

How to run it

Prerequisites – Docker Compose and an OpenAI‑compatible LLM API key (e.g., DeepSeek, Qwen, Zhipu).

Setup

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

Visit http://localhost:3000 for the front‑end and /admin for the admin UI (password is in the generated .env). After a few minutes the first batch of items will be processed.
For production you’ll need a domain, HTTPS, and (if deploying in mainland China) a Chinese‑friendly server – see docs/deploy.md.

Customising for your own industry

All industry‑specific data lives under the industry/ folder:

site.ts – site name, tagline, industry keywords.
taxonomy.ts / topics.json – categories, tags, topic pages.
sources.json – the initial list of feeds (replace the demo AI‑news sources with your own).
prompts/ – the exact LLM prompts used for filtering, scoring, and writing. Edit these to encode your domain knowledge.
selection.ts – numeric thresholds for the dual‑score stage.
features.ts – toggle AI‑only modules (model leaderboard, Codex monitor).
brand/ & pages/ – logos, usage policies, privacy text.

The README suggests using an LLM‑assistant (Claude Code, Codex, etc.) to read AGENTS.md and docs/customize.md and automatically rewrite the files for a target sector such as law, HR, finance, etc.

Documentation highlights

把它改成你的行业 – step‑by‑step guide to re‑brand and re‑configure.
信源 – details on the six supported source types and how to add new ones.
精选与校准 – explains the dual‑scoring process and how to calibrate it with your own labelled data.
部署 – production‑grade Docker, domain/HTTPS setup, cost considerations.
架构 – three processes (crawler, worker, web server), immutable rules, directory layout.
模型榜与 Codex 重置监控 – AI‑specific dashboards.

Who might use this?

Media companies wanting a fast, AI‑augmented news‑digest for a niche vertical.
Enterprises that need an internal “industry pulse” site (e.g., legal updates, HR policy changes, commodity price news).
Researchers or hobbyists looking for a ready‑made pipeline to experiment with LLM‑based summarisation and clustering.

License & branding

The code is released under the MIT License. The name AIHOT and its logo are not covered by the license; you must replace them if you redistribute the product.

In short: AIHOT is a complete, Docker‑ready stack that turns any collection of web feeds into an AI‑curated, Chinese‑language news portal with daily briefings, clustering, and hot‑topic ranking. All the AI‑specific logic (prompts, thresholds, clustering rules) is exposed for easy adaptation to any industry.