LLM Wiki Architecture

Local-first knowledge base built on the Karpathy-style markdown wiki pattern, maintained by Codex agents, captured from Obsidian Web Clipper, versioned on GitHub, and published through VitePress plus Vercel for browser access on desktop and mobile.

web clipping manual notes screenshots / OCR new raw files ingest / lint / query wiki synthesis scheduled runs git push deploy trigger site publish browse static pages vault sync watch folder / cron handoff local repo remains source of truth until pushed Capture Inputs Chrome Obsidian Web Clipper page -> md User manual notes local edits new ideas Images screenshots attachments OCR support Local Knowledge Workspace llm-wiki repo raw/ immutable-ish capture layer wiki/ maintained synthesis layer Karpathy-style markdown knowledge base Obsidian local vault browsing search, backlinks, quick edits desktop authoring surface Bilingual corpus English canonical pages Chinese .zh siblings durable markdown knowledge Cron / scheduler periodic tidy-up prompts ingest backlog and lint passes optional local automation hands work to Codex agent sessions Ingest loop read raw/*.md extract OCR if needed update wiki/sources touch related topics new material integration Codex agent reads AGENTS.md checks index/log edits and validates commits intentionally maintenance control plane Wiki outputs wiki/sources wiki/topics wiki/concepts wiki/answers persistent synthesis Automation boundary Versioning and Publish Pipeline Git workflow small intentional commits repo-local identity local change packaging GitHub origin source of truth history, sync, sharing collaboration remote VitePress site markdown viewer repo layout preserved index + docs + raw + wiki browser-readable presentation Hosted Access Layer Vercel binds GitHub repo auto deploy on push free hosted deployment Domain endpoint vercel.app or custom host public read access single URL for all devices Desktop / Mobile browser reading phone and laptop no local vault needed Published knowledge static, searchable, shareable markdown Maintenance principles ingest, query, lint, bilingual sync, git discipline Legend local app / viewing surface agent / workflow knowledge artifact cloud / hosting ingest / publish event scheduled automation

Local-First Capture

  • • Chrome + Obsidian Web Clipper turns web pages into markdown under raw/.
  • • Obsidian stays the fastest local reading and editing surface for the vault.
  • • Screenshots can join the corpus with OCR-backed evidence before synthesis.

Codex Maintenance Loop

  • • Codex agents read repo rules, ingest new raw sources, and repair cross-links.
  • • Scheduled runs can proactively perform lint, backlog cleanup, and wiki updates.
  • • The durable result remains markdown, bilingual, and git-versioned.

Publish Everywhere

  • • Pushes to GitHub trigger Vercel deployments for the VitePress site.
  • • The site exposes the same wiki from a free hosted domain or custom domain.
  • • Any desktop or phone browser can read the knowledge base without Obsidian.

LLM Wiki • generated for repository architecture documentation • updated 2026-04-17 13:51 Asia/Shanghai