Deploy Hf Spaces Vercel

• Expose port 7860 (HF default).
• Non-root user.
• Pre-download the ONNX model at build time via huggingface_hub.snapshot_download so cold start is < 5s.

Web → Vercel

• Connect GitHub repo.
• Root directory: apps/web.
• Framework preset: Next.js.
• Install command: pnpm install --frozen-lockfile.
• Build command: pnpm build.
• Env vars: NEXT_PUBLIC_API_URL=https://<user>-<space>.hf.space.

Postgres → Neon

• Free tier: 0.5 GB storage, branching, auto-suspend.
• Enable pgvector extension: CREATE EXTENSION IF NOT EXISTS vector;.
• Use the pooled connection URL for the API (...-pooler.neon.tech).
• Store DB URL in HF Space secrets + Vercel env vars (web doesn't hit DB directly; for migrations in CI).

Redis → Upstash

• Free tier: 10K commands/day, sufficient for demo.
• Use REST-compatible URL for arq — or direct TCP if preferred.

Model weights → HF Hub

• Push ONNX artifacts to a model repo: clinical-nlp/ner-v1.
• Tag production for the currently-deployed version.
• API loads via huggingface_hub.hf_hub_download at startup (cached on Space persistent storage).

Deploy order (the /ship command enforces)

1. Run DB migrations first (GitHub Actions job migrate against Neon).
2. Deploy API to Space.
3. Smoke test: curl https://<space>.hf.space/health → 200.
4. Deploy web to Vercel (auto on main push).
5. Run Playwright smoke against production URL.
6. Tag release in git.

Rollback

• Space: revert to previous Space commit (HF keeps history).
• Vercel: one-click rollback in dashboard.
• DB: migrations must be reversible (see sqlalchemy-async skill).
• Model: change HF Hub production tag back to previous version, redeploy API.

What's banned

• Pushing untested code to the Space main branch.
• Skipping the smoke test between API deploy and web deploy.
• Hardcoding secrets (anywhere — ever).