◉ HASSELT · BEBook →
← CASE STUDIES·RESTAURANT · HASSELT, BE

Drugstore Grand Café answers every guest in four languages.

A multilingual RAG chatbot grounded in the café's own menu, allergen tables, and reservation documentation. No hallucinations by design. Deployed in 14 days.

~90%
INQUIRIES AUTO-RESOLVED
4
LANGUAGES SUPPORTED
14
DAYS TO PRODUCTION
/ THE SITUATION

Repetitive guest questions were interrupting service, across four languages.

Guests asked the same questions daily: allergen information, menu details, reservation policy, opening hours, event hire availability. Staff handled each one manually, often mid-service. In a Belgian city where guests switch between Dutch, French, English, and German in the same week, every inquiry required a different response. The answer was always in the documentation; the bottleneck was finding it fast enough.

/ WHAT WAS BUILT

A RAG assistant that knows the café better than anyone.

Every answer grounded in Drugstore's own indexed content, menu, allergens, specials, reservations, events. If the answer is not in the docs, the chatbot says so and offers to connect the guest to a human.

01

Documentation indexed

Full menu, allergen tables, daily specials format, reservation policy, event hire packages, gift voucher terms, and opening hours, all ingested into Pinecone, chunked and embedded via Cohere.

02

Retrieval pipeline calibrated

Top-K retrieval tuned against 200 real guest questions sampled from the enquiry inbox. Hallucination rate validated at zero on the test set before going live.

03

Language layer added

Automatic language detection from the first message, no guest setting required. DeepL handles edge cases where Groq confidence drops below threshold. NL, FR, EN, and DE supported from day one.

04

Deployed in 14 days

Single script embed on the existing website. Human handoff connected to the reservations inbox. Monitoring and weekly quality reviews in place from day one.

05

Monitoring dashboard set up

Every conversation logged and tagged by resolution outcome. Weekly review identifies low-confidence responses for improvement before they become a recurring issue.

06

Continuous improvement loop

New menu items, seasonal specials, and event packages pushed to the vector store via a simple upload, no developer involvement. The chatbot stays current with the kitchen.

/ RESULTS

Staff answer fewer questions. Guests wait less.

~90%

Auto-resolve rate

Of all incoming guest inquiries answered by the chatbot without human involvement, including after-hours.

4

Languages supported

NL, FR, EN, and DE. Language detected automatically from the first message, no guest setting required.

0

Hallucinations

RAG architecture grounds every answer in the café's indexed documentation. Ungrounded answers are refused.

14

Days to production

From signed contract to a live, embedded chatbot on the Drugstore website.

/ STACK

The smallest stack that solved the problem.

PineconeCohere EmbeddingsGroq InferenceDeepLNext.jsVercel
/ YOUR SYSTEM

The same system
on your docs.

30-minute call. We'll index a sample of your documentation and show you a live prototype, free, no deck.