vraagamsterdam
Transparency

How we work. No hidden agenda.

Compliance is not an afterthought for us. This is what we do, which models we run, and what happens to your questions.

Which models

We run Mistral 7B Instruct via Ollama in development and the Mistral API (France, EU-hosted) in production. Embeddings via bge-m3: multilingual, open source, free. No OpenAI, no Anthropic, no Google.

Which data

Only publicly available sources: amsterdam.nl, rijksoverheid.nl, belastingdienst.nl, and a manually curated set we re-evaluate every quarter. Never personal data, never DigiD, never BSN.

What we log

  • Question, answer, sources, model version, prompt version, latency. All anonymized.
  • No IP addresses after the initial rate-limit check.
  • No tracking cookies, no third-party analytics tags.
  • Audit log retained for 7 years per the Dutch Archives Act.

EU AI Act

vraag.amsterdam is a limited-risk AI system under the EU AI Act. We publish a quarterly transparency report with accuracy metrics, crisis-detection false-positive rate, and confidence distribution. FRIA prepared at go-live.