Vale construir sua versão de Littlebird?

Build a local personal AI assistant web app in an empty folder. No accounts, no cloud, no telemetry.

Stack, no substitutions:
- TypeScript, Node 20, Fastify for the server
- SQLite via better-sqlite3, file at ./data/assistant.db
- Server-rendered HTML with htmx for interactivity, plain CSS, no React, no Tailwind
- OpenAI-compatible chat completions and embeddings, base URL and key read from .env (OPENAI_API_KEY, OPENAI_BASE_URL, CHAT_MODEL, EMBED_MODEL)
- node-cron for scheduled jobs, all in one process

Features in scope:
1. Context library: paste text, upload .md/.txt, or submit a URL. URLs are fetched server side and reduced to readable text with @mozilla/readability plus jsdom. Chunk to ~800 tokens, embed, store chunks and vectors in SQLite.
2. Chat: a single conversation view. On each message, retrieve the top 8 chunks by cosine similarity, put them in the system prompt with source titles, stream the reply back. Persist all turns.
3. Watchlist: each row is a name, a URL or a search query, and a cadence (daily or weekly). A cron job fetches each item, diffs the extracted text against the last stored snapshot, and asks the model to summarize what changed in under 80 words. Skip items with no meaningful diff.
4. Digest: a /digest page listing the latest summaries newest first, plus a digest.md file written to ./out/ after each run so it is greppable.
5. Every model call is logged to a calls table with model, token counts and latency, and /usage shows totals by day so you can see what you are spending.

Out of scope, do not build: user accounts, OAuth to any third party, email or push notifications, mobile app, multi-user sharing, vector database services, Docker.

Deliver: npm install, npm run dev on port 3000, a seeded example watchlist item, .env.example with every variable, and a README that states plainly what works and what breaks if the process dies.