What AIHOT Is and the Problem It Solves
AIHOT is an open-source framework for building self-hosted news aggregation sites that use large language models to discover, filter, and summarize industry-specific hot topics. The project originated as a personal tool for tracking AI industry developments but was released publicly after receiving requests from professionals in fields like law, HR, finance, and precious metals who wanted similar tools for their own domains. The creator notes they lack the domain expertise to build such tools for every industry, so open-sourcing the framework allows others to adapt it using their own knowledge of relevant sources and criteria for what constitutes a “hot” topic in their field.
The core problem AIHOT addresses is the manual effort required to monitor multiple information streams, identify meaningful developments amid noise, and produce timely summaries. By automating initial filtering, deduplication, and summarization while allowing human oversight and customization, the framework aims to reduce the labor of staying informed in fast-moving industries.
Why It Is Gaining Traction
The project has attracted attention from developers and industry professionals seeking customizable solutions for internal knowledge monitoring or public-facing industry news sites. Its appeal lies in the transparency of its prompt engineering and scoring logic — all selection criteria and model prompts are publicly available in the repository, enabling users to adapt the system to their specific needs without modifying core code. This approach lowers the barrier to entry for domain experts who understand what matters in their field but may not have deep technical expertise.
With over 5,400 stars on GitHub, the project indicates growing interest in self-hosted, AI-augmented information tools that prioritize user control over data and customization. The emphasis on replacing default sources and prompts with industry-specific equivalents reflects a broader trend toward adaptable AI infrastructure rather than one-size-fits-all SaaS offerings.
How It Works at a High Level
AIHOT processes incoming content through a series of automated steps powered by large language models. Content is first gathered from configured sources, deduplicated, and subjected to an initial screening. Items deemed potentially important are then scored independently twice using LLMs, with prompts and thresholds defined in human-editable files. Content that meets the scoring threshold and is not a duplicate of already-selected items proceeds to the curation stage.
Selected articles are clustered into events based on semantic similarity — initially using text overlap or vector embeddings if a vector service is configured — followed by model-based verification to determine whether reports describe the same event, a follow-up, or unrelated topics. Confident merges are finalized; uncertain cases are re-evaluated with a different model. This ensures that multiple reports on the same development are consolidated into a single event entry.
Hotness is calculated per event, not per article, with weighting designed to prevent inflation from repetitive reporting: each independent source contributes only once within a 48-hour window, with influence halving after 24 hours. This means a story gaining traction across diverse outlets ranks higher than one heavily repeated by a single publisher.
The system generates daily digests at 08:00 local time, summarizing selected events without re-running models on already-processed content. Weekly and monthly reports are compiled from these digests, with models only contributing introductory summaries and section guides. Outputs are available via RSS, public APIs, MCP endpoints, and Markdown formats designed for both human readers and AI agents.
Who It Is For and Real Use Cases
AIHOT is intended for individuals or teams who want to create a tailored news or intelligence site for a specific industry, profession, or interest area. The documentation emphasizes that users must supply their own sources and refine the selection criteria — referred to as “KnowHow” — to reflect what matters in their domain. Example use cases include tracking regulatory changes in finance, monitoring clinical trial updates in healthcare, or following policy developments in environmental law.
The framework supports six source types: RSS feeds, web page lists, JSON endpoints, X (Twitter) accounts, WeChat public accounts, and custom scripts for pushing content. Sources can be tagged by tier (e.g., official vs. media) to influence scoring thresholds. After deployment, users can manage sources, test content ingestion, and diagnose issues through an admin interface accessible at /admin.
For integration with AI agents or workflow automation, AIHOT provides structured data endpoints, agent-friendly Markdown feeds, and compatibility with protocols like MCP. The project’s architecture uses Node.js 24, TypeScript, React Router with server-side rendering, Fastify, PostgreSQL, pg-boss for job queues, Tailwind CSS, and Docker Compose for deployment.
Honest Caveats and Considerations
The creator explicitly states they are not a professional developer and that the code, while improved through AI-assisted rewrites, may still contain areas of suboptimal quality. Users are encouraged to report issues via GitHub, though responses may be delayed. The repository contains the same engine code used to power the live AIHOT site, but it is not presented as a polished, generic framework — updates from the live site are synchronized when possible but not guaranteed.
Critically, the repository does not include the live AIHOT site’s source list or operational data. It ships with 18 publicly available international AI news sources as a demonstration set, sufficient to verify functionality but requiring replacement with industry-specific sources for meaningful use. Users are advised not to reuse the AIHOT name or logo and to rebrand the site with their own identity.
Customization primarily involves editing files in the site/ and industry/ directories, including site.ts for branding, taxonomy and topic files for classification, sources.json for initial feed configuration, prompts/ for selection and writing standards, selection.ts for thresholds, and models.ts for assigning LLMs to different pipeline steps. The documentation recommends validating changes using a personal set of annotated examples and the provided evaluation script (scripts/eval-selection.ts) to measure precision and recall before deploying.
Deployment requires Docker and an OpenAI-compatible API key (with examples given such as DeepSeek, Qianwen, or Zhipu). Initial content processing after startup may take approximately thirty minutes to complete. Guidance is provided for environments without Node.js, servers in mainland China, and configurations involving custom domains and HTTPS.
While the project enables powerful automation, its effectiveness depends entirely on the quality and relevance of user-provided sources and the appropriateness of the customized scoring logic. There are no guarantees of performance, accuracy, or suitability for any particular use case, and the framework should be evaluated in context rather than assumed to deliver journalistic or analytical rigor out of the box.
Source: https://github.com/KKKKhazix/AIHOT
