Yoveletta – Swiss Cloud POS with a Custom AI Assistant

Full web relaunch for Yoveletta: Next.js 16, React 19, Tailwind v4 — combined with an in-house Llama 3.3 70B RAG chatbot trained on 17 Yoveletta PDFs. Real-time streaming, fully hosted in Switzerland.

Yoveletta doesn't sell an off-the-shelf POS — it sells an industry-specific solution for hospitality, retail and services, from food trucks to bakeries to bars. That depth was invisible on the old site: an outdated WordPress presence, four product packages crammed into a single pricing table, no English version, and no technical support outside office hours. At the same time, the relaunch couldn't lose any rankings or backlinks — and the entire stack had to run on Swiss infrastructure, with zero customer data leaving Switzerland for US cloud providers.

We rebuilt Yoveletta.ch from scratch: Next.js 16 (App Router) with React 19 and Tailwind v4, paired with a custom Node server in front of Vercel Edge and a FastAPI function for the AI answers. Over 60 pages in DE and EN — every industry, every module, every pricing tier maintained twice and cleanly separated via the /en/* URL prefix. The core is a RAG chatbot that exclusively answers from 17 Yoveletta PDFs: terms of service, setup guides, receipt-printer configuration, card-terminal integration, refund process, the Yoveletta brochure. Each chunk is vectorised through the Supabase Edge function 'embed' using gte-small (384 dimensions) and stored in an HNSW index. On each query we fetch the top-6 chunks via pgvector cosine similarity and send them along with a strict system prompt to Llama 3.3 70B on Groq. Answers stream token by token in real time. If the model invents anything that isn't in the retrieved context, it admits so — and points the user to verkauf@yoveletta.ch. On top of that: a 3-step booking modal with a custom date picker and slot logic, a Bento hero with a GSAP animation, full Google Consent Mode v2 with default-deny and active cookie cleanup, three JSON-LD schemas (Organization, SoftwareApplication, LocalBusiness), and nine permanent 301 redirects that migrate the entire WordPress link equity.

• Next.js 16 (App Router)

  React 19, Server Components, Standalone-Build
• Tailwind CSS v4

  @theme Design Tokens, fluide Typografie via clamp()
• FastAPI · Python 3.13

  Streaming-Endpoint für den RAG-Chatbot
• Supabase + pgvector

  Postgres-Vector-Store mit HNSW-Index
• Llama 3.3 70B · Groq

  Streaming-LLM für token-by-token Antworten
• gte-small Embeddings

  384-dim, in Supabase Edge-Function gehostet
• GSAP 3.15

  Bento-Hero-Timeline, Hover-Animationen
• Vercel Edge

  Hybrid Next.js + Python Serverless

Yoveletta now has a website that matches its own depth: faster, modular, bilingual — and built on a stack that only became production-ready in 2026. The AI assistant answers 24/7 the questions that used to land in the support inbox, without ever hallucinating. Sales inquiries now arrive through the booking modal with pre-qualified time slots — no more drawn-out email ping-pong. All 17 internal documents became a knowledge base, without a single character leaking to OpenAI.