Hüni — KI-Gemeindeassistent
KI-Gemeindeassistent für Hünenberg. RAG-basierter Chat mit Halluzinations-Stopp, Quellenangaben und Schweizer Datenhoheit.
Problem
Citizens lose time searching for official information on deeply nested municipal websites. Hünenberg already operates a modern online services portal and the eZug app, yet the most common queries (change of address, dog registration, waste calendar) still lead to phone calls or counter visits.
Commercial chatbot solutions are unsuitable for a Swiss municipality: they hallucinate when knowledge is missing, process data outside Switzerland, and provide no reliable source citations. Static FAQs, in turn, do not scale to the breadth of real citizen questions.
Solution
Hüni is an AI municipal assistant based on Retrieval-Augmented Generation (RAG). The knowledge base consists exclusively of verified municipal sources: regulations, sitemap, online services portal, waste calendar.
A strict system prompt enforces three non-negotiable rules: source citations on every substantive answer, active refusal to process personal data, and a hard hallucination guardrail — when the knowledge base does not answer a query, the response consistently escalates to the municipal office.
Architecture: hosting entirely in Switzerland, FADP-compliant data processing. Estimated operating costs for a municipality of 9,000 residents: CHF 200–400 per month.
Target audience
Primary:
- Mid-sized Swiss municipalities (5,000–25,000 residents)
- Cantonal administrations focused on citizen services
- Municipal offices with high volumes of recurring queries
Secondary:
- Municipalities with high shares of international residents (multilingual needs)
- Municipal associations building shared-service models
- GovTech vendors as white-label component
Tech Stack
Compliance
Roadmap
Q2 2026 — Pilot
Functional prototype with curated knowledge base (~30 entries), browser-based widget and offline mirror. Hallucination tests against edge cases. Stakeholder demo to Hünenberg municipal council.
Q3 2026 — Backend
Cloudflare Worker as API proxy. Vector DB (Qdrant, Swiss hosting). Sitemap crawler for automated ingest. Friendly-user test with 50 Hünenberg residents.
Q4 2026 — Go Live
Production launch on huenenberg.ch. Analytics dashboard for administration (most frequent queries, escalation rates, usage patterns). Iterative knowledge base refinement.
2027 — Scaling
Adoption by 2–3 additional Zug municipalities. Full multilingual support (FR / IT / EN). eZug-eID integration for personal, authenticated queries.
References & Reading
- Lewis, P. et al. (2020): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401.
- Swiss Federal Data Protection and Information Commissioner (FDPIC): Guidelines on the use of AI in public administration, 2024.
- Federal Chancellery: Digital Switzerland Strategy 2024–2027.
- Canton of Zug: eZug — Digital Administration 2026, Strategy Paper.
- OECD (2023): Recommendation on Artificial Intelligence — Public Sector Application.
Screenshots
FAQ
How does Hüni prevent hallucinations?
Are personal data processed?
Where is the data hosted?
How current is the knowledge base?
What does it cost to run?
Can other municipalities adopt Hüni?
What happens with legally binding queries?
Which language model is used?
Case Study
A Swiss municipality of 9,000 residents tests an AI assistant that makes citizen services available 24/7 — without the risks of commercial chatbots.
Hünenberg is among the most digitally advanced municipalities in the Canton of Zug. With the eZug app, an established online services portal, and an active Energy City Gold strategy, basic digitalisation is long since completed. The next question: how can AI tools deepen citizen services without compromising cantonal standards of data sovereignty and legal certainty?
The Hypothesis
A tightly curated RAG bot, fed exclusively from official municipal documents and equipped with a hard hallucination guardrail, could answer the most common citizen queries — change of address, dog registration, waste calendar — without human effort while cleanly escalating complex requests to administration staff.
The Prototype
A fully functional prototype with 30 knowledge base entries, a browser-based widget, and explicit hallucination tests against typical consumer-AI weaknesses ("What are next Saturday's lottery numbers?") was built in 14 days. The system strictly enforces source citations and consistently refers to the municipal office when no official source is available.
The Path to Production
The pilot is designed as a modular architecture: a sitemap crawler replaces the manual knowledge base, an API proxy solves security, the vector database preserves data sovereignty. All data in Switzerland, FADP-compliant. Estimated operating costs: CHF 200–400 monthly for a municipality of 9,000.
The key difference from Chatbase and commercial solutions: Hüni is not allowed to improvise. What is not in the source is not said.
More projects
Clarivio
Dashboard prototype for the visual representation of cognitive diagnostics, biomarker tracking and clinical workflows. React-based, visually guided by publications such as SMC 2024 and NIA-AA 2024.
epiLoom
Prototype of an ILAE-2017-oriented epilepsy monitoring platform, conceived as a proof of concept for Swiss neurology departments. The project demonstrates real-time patient monitoring, AI-assisted seizure prediction and alignment with Swiss medical standards (ICD-10-GM, Swissmedic, EPD).
Arcanum
A zero-knowledge password manager with vault functionality and AI-assisted risk analysis. Developed as an alternative to existing solutions — with the ambition of a more stable architecture, stricter encryption and more intelligent usability.
Sonder
Rethinking artist discovery: curated setcards for independent musicians — without algorithms, charts, or like-driven visibility.
What should we build together?
This project is part of our lab research. If you need something similar — your own AI tool, dashboard, custom workflow — talk to us. Every engagement starts with an honest assessment, not a sales pitch.