Hi-Lite × Sia Huat Product Assistant

Phase 1 is a conversational product-enquiry web application using Claude Sonnet 5 for replies and image analysis, OpenAI gpt-transcribe for voice transcription, and a Sia Huat catalogue in Supabase Postgres. The web application's runtime does not call the former n8n workflows.

See docs/CLAUDE-RESTORATION.md for the current provider configuration and verification.

Live demo

Open the Hi-Lite × Sia Huat Product Assistant

Current release: V_3.1.2

The website catalogue is discovered from Sia Huat's public sitemap and imported with the displayed item code as the unique identifier. Product discovery uses the cached Supabase catalogue; after the customer confirms an exact item, the app performs a fresh website check for the displayed price and stock status.

The assistant clarifies broad requests, searches real catalogue data, asks for quantity, calculates an estimated total deterministically, and prepares the result for human sales review. It never confirms an order or sends a customer message automatically.

See docs/ARCHITECTURE.md for the system design, trust boundaries, security model, and remaining production work.

Local setup

1. Copy .env.example to .env.local and add the server-only Anthropic and OpenAI keys and existing catalogue configuration.
2. Install dependencies with pnpm install.
3. Run pnpm dev.
4. Open http://localhost:3000 or the port printed by Next.js.

Required application variables:

ANTHROPIC_API_KEY=your-anthropic-key
ANTHROPIC_MODEL=claude-sonnet-5
OPENAI_API_KEY=your-openai-key
OPENAI_TRANSCRIPTION_MODEL=gpt-transcribe
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_PUBLISHABLE_KEY=your-publishable-key

The catalogue import scripts additionally require SIA_HUAT_PROJECT_REF and SIA_HUAT_POSTGRES_PASSWORD. Those administrative values are for local import/migration work and should not be configured in the deployed web application.

Catalogue sync

The public sitemap currently exposes more than 19,000 product pages. The crawler is resumable and writes sorted results plus a failure report under tmp/siahuat-crawl/.

pnpm catalogue:crawl -- --concurrency=8
pnpm catalogue:crawl -- --concurrency=8 --import

Use a small --limit=10 while testing. The checked-in n8n workflow export is n8n/sia-huat-b2b-phase-1.json; import it into n8n and configure SUPABASE_URL and SUPABASE_PUBLISHABLE_KEY in that n8n environment.

Verification

pnpm typecheck
pnpm lint
pnpm build
pnpm db:verify
pnpm qa:all

pnpm qa:text runs the Sia Huat conversation, relevance, memory, safety and handoff regression suite derived from the chatbot SOP. pnpm qa:image checks the two supplied product-photo fixtures. Both commands write machine-readable evidence to tmp/qa-reports/.

The local test flow should cover:

1. Ask for a broad product such as a knife.
2. Answer the clarification.
3. Confirm that displayed products match Supabase.
4. Select a product and enter a quantity.
5. Verify the calculated total and human-review wording.
6. Upload a product photo. Claude inspects the actual pixels; catalogue lookups verify any potential match. Filenames are never visual evidence.

Deployment policy

Development and acceptance testing stay local. GitHub is the source repository. Vercel should be connected and deployed only once the user explicitly approves a release candidate, avoiding unnecessary builds during development.