Browser-Native AI Knowledge Site

A personal website that behaves like queryable documentation. Visitors ask natural-language questions and get answers grounded in published Markdown — with source citations. All retrieval and AI inference happen in the visitor's browser. No backend, no inference API, no per-answer cost.

If a browser can't run the on-device model, it gracefully falls back to local search, and can hand the grounded prompt off to ChatGPT.

How it works

Markdown (src/content) ──build──▶ chunks + search index (public/search-index.json)
                                        │
                        Browser ◀───────┘
                          │
             MiniSearch retrieval ─▶ top chunks ─▶ WebLLM (WebGPU) ─▶ answer + Sources
                                              └─ (no WebGPU) ─▶ top matches only
                                              └─ Continue in ChatGPT (copies grounded prompt)

• Content — every fact lives in src/content/**/*.md with frontmatter (title, description, tags). This is the single source of truth.
• Build-time index — scripts/build-index.mjs splits each doc by heading into chunks and writes public/search-index.json, plus llms.txt / llms-full.txt for external AI systems. Runs automatically via prebuild.
• Retrieval — src/components/ask/search.ts loads the index into MiniSearch and returns the top-N chunks entirely client-side.
• Inference — src/components/ask/webllm.ts runs a ~1B model (Llama-3.2-1B-Instruct) via WebLLM/WebGPU. Downloaded only after the visitor clicks Enable Local AI, then cached for near-instant repeat visits.
• Grounding — src/components/ask/prompt.ts builds the same context-constrained prompt for both the local model and the ChatGPT handoff: only answer from context, never invent, cite sources.

Develop

pnpm install
pnpm dev      # rebuilds the index, then starts Astro at http://localhost:4321
pnpm build    # static output in dist/ (prebuild regenerates the index)
pnpm preview  # serve the production build locally

Deploy dist/ to any static host (Cloudflare Pages, Vercel, Netlify, GitHub Pages). No server configuration required.

Editing the knowledge base

Add or edit files in src/content/. Each heading (## / ###) becomes a retrievable, individually-citable chunk. Rebuild (pnpm build) and the index, pages, and llms.txt regenerate automatically.

Stack

Astro · React · TypeScript · MiniSearch · WebLLM · 100% static.