AI Chatbot Demos
Two working websites where an AI assistant does real work — it searches the site's own data, fills a cart, books an appointment — instead of just chatting.
Italiano
Both sites are fictional and open to everyone. They share one architecture and do two very different jobs:
| Demo | What the assistant does | Live |
|---|
Crema & Ferro (web/) | A coffee-gear shop: recommends and compares products, adds them to the cart, tracks orders, answers from the help center. Includes a step-by-step tutorial of the code. | https://chat-bot.giuseppebosi.com |
Poliambulatorio Vallenova (clinica/) | A medical clinic: finds services and doctors, computes free slots, books and cancels only after the patient clicks Confirm, and refuses to give medical advice. | https://chatbot-clinica.giuseppebosi.com |
The sites, their content and some internal notes are in Italian; this README is the English entry point.
What is inside
- Next.js 16 (App Router), React 19, Tailwind CSS, shadcn/ui, AI Elements
- Vercel AI SDK 7: streaming chat, typed tools, user approval for every action that changes data
- OpenRouter free models: a preferred model, fallbacks, and a stream guard that never shows an empty reply or the model "thinking out loud"
- PostgreSQL + pgvector with Drizzle ORM: hybrid search (full text + vectors, fused with RRF)
- Local embeddings (multilingual-e5-small via Transformers.js): no external embedding service
- An embeddable widget (
<script> + lazy iframe) for third-party sites
- Abuse protection: per-visitor and site-wide rate limits, public widget key, origin and content-type checks, body size limits
- Playwright end-to-end tests and Vitest unit tests
- Deployment with Docker Compose behind Caddy:
deploy.sh and rollback.sh
- Search and AI friendliness:
sitemap.xml, robots.txt, llms.txt, Open Graph, JSON-LD
Run it
Requirements: Docker and a free OpenRouter API key.
cd web # or: cd clinica
# create .env.production with:
# OPENROUTER_API_KEY=... POSTGRES_PASSWORD=...
# TOOL_APPROVAL_SECRET=... IP_HASH_SALT=...
./deploy.sh # builds the image, starts app + Postgres, waits for /api/health
Migrations and seed data run when the app starts. The app listens on 127.0.0.1 only: put a reverse proxy (Caddy) in front of it. ./rollback.sh returns to the previous image without touching the data. Each app's README.md lists the development commands.
Structure
web/ Crema & Ferro shop + tutorial (Next.js app, own database and compose file)
clinica/ Vallenova clinic (same stack, different tools and rules)
.claude/ Agent setup used to build the project: skills, reviewers, rules
docs/ Research notes written before each technical choice
License
- Texts and content (pages, catalog, clinic information, tutorial): CC BY 4.0
- Code: MIT
Names, people, prices and orders are invented; any resemblance to real businesses or people is coincidental.
Author
Giuseppe Bosi — Master's-level degree in Electronic Engineering (5-year Italian Laurea, EQF level 7); teacher of Computer Science and Technology in Italian secondary schools.
www.giuseppebosi.com · LinkedIn
