Architecture

中文 | English

This page summarizes the current llm-wiki operating model as a local-first knowledge system with an automated maintenance loop and a static publishing pipeline.

Open the standalone architecture diagram

Summary

• Capture starts locally: Chrome with Obsidian Web Clipper converts web pages into markdown and stores them under raw/.
• The repository follows the Karpathy-style LLM knowledge base pattern: raw/ is the source layer and wiki/ is the maintained synthesis layer.
• Codex agents act as the maintenance engine: they ingest new sources, update durable wiki pages, run lint passes, preserve bilingual siblings, and commit the resulting changes.
• GitHub is the canonical remote. After local commits, pushes propagate the updated vault to the shared source of truth.
• VitePress renders the repository into a readable site, and Vercel deploys it automatically from GitHub so the wiki is available from desktop and mobile browsers.

Flow

1. Web content or notes enter the vault through raw/.
2. Codex maintenance sessions integrate the new material into wiki/sources/, wiki/topics/, wiki/concepts/, and wiki/answers/.
3. The repository history remains intentional and inspectable through Git.
4. GitHub push triggers Vercel deployment of the VitePress site.
5. Readers access the same knowledge base through a public URL in any browser.

Design Goal

The core idea is to keep the knowledge base editable as plain markdown locally, while making maintenance increasingly automatic and publishing essentially zero-touch after each push.