ChapterChat algorithm

Province of Groningen

The primary aim is to give citizens who qualify for specific municipal schemes, such as "Insulation and Nij Begun" (Measure 29), easy insight into their rights and obligations without having to decipher legal texts themselves.

Last change on 7th of July 2025, at 10:57 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: In development

General information

Theme

• Living
• Nature and Environment
• Space and Infrastructure
• Health and Healthcare

Begin date

07-2025

Contact information

[e-mail]

Link to publication website

www.provinciegroningen.nl

Responsible use

Goal and impact

The primary aim is to give citizens who qualify for specific municipal schemes, such as "Insulation and Nij Begun" (Measure 29), easy insight into their rights and obligations without having to decipher legal texts themselves.

Considerations

The system never communicates directly with citizens. No communication to citizens takes place without prior review and possible adjustment by the knowledge staff of the province of Groningen.

Human intervention

No communication to citizens will take place without prior review and possible adjustment by the knowledge staff of the province of Groningen.

Risk management

There is no risk of automated decision-making or communication and the algorithm has no impact on fundamental rights because the algorithm does not make decisions with legal consequences. It only proposes an email to grant applicants. All proposals are reviewed by a concerned employee of the province of Groningen.

Legal basis

Measure 29: insulation approach for houses in Groningen and North Drenthe

Nij Begun is a foundation of the Province of Groningen and municipalities in Groningen and North Drenthe. With Nij Begun, the government takes 50 measures to support inhabitants of the earthquake area in Groningen and North Drenthe. Measure 29 is one of those measures. This rule is officially called 'The subsidy scheme for insulation and ventilation of buildings, houseboats and caravans province of Groningen and the municipalities of Aa and Hunze, Noordenveld and Tynaarlo'. The subsidy scheme is for insulating and improving the ventilation of houses in the earthquake area. Thus, these reach the standard for home insulation. Samenwerkingsverband Noord Nederland (SNN) implements Measure 29. Residents therefore apply for the subsidy on SNN's website. We share data with SNN on previous subsidy applications to RVO.

The processing of personal data within the chatbot and the underlying platform takes place on legal bases from the General Data Protection Regulation (AVG/GDPR) including:

• Using the disclaimer(s), we clearly indicate the purpose connection, do minimal data processing sufficient to what is necessary (contacting), ensure correct information sources underlying the chatbot, anonymise all traceable PII-related data points with exclusion of contact data where consent is given, limit storage to 30 days (Art. 5 para 1a to f AVG).
• The processing of service-desk messages and other human contact moments requires explicit user consent (Art. 6 para 1a AVG).
• For logging and analysis of anonymous session data for the purpose of quality assurance and improvement, legitimate interest is used (art. 6 subsection 1f AVG).

Links to legal bases

Maatregel 29: isolatieaanpak voor woningen in Groningen en Noord-Drenthe: https://www.rvo.nl/onderwerpen/maatregel-29

Elaboration on impact assessments

No formal DPIA or AI impact assessment has been carried out.

Operations

Data

The Chapter algorithm primarily uses verified legal texts, approved by Nij Begun and Milieu Centraal, as a source of knowledge in the vector store. Personal data is only processed after explicit consent, for example when filling in a service desk form. This includes: name, e-mail address, subject, message, IP address, browser data, date, and time. For quality control purposes, anonymised session data (such as hashed user IDs) are temporarily stored. All personal data will be deleted or anonymised after a maximum of 30 days.

Technical design

The Chapter algorithm runs in a defined cloud environment and operates according to a Retrieval-Augmented Generation (RAG) architecture. The knowledge base consists of one or more vector stores containing only legal texts verified by Nij Begun and Milieu Centraal. After an initial guardrail check, each user query is processed via specialised search paths by various retrievers (including RAPTOR, Hybrid, MMR and AgenticRAG with Hypothetical Document Embeddings and Self-Query Refinement). These retrievers select the most relevant passages, after which the final response is generated using advanced Large Language Models, including GPT-4.1, Mistral Large, Claude 4.0 Sonnet, Gemini-pro 2.5, or, for example, Grok 3.

The generated response is presented to the user via TLS-encrypted API endpoints via a web application running on Azure containers. The infrastructure is scalable: at least one worker is always active and additional workers are automatically added as traffic increases, up to a maximum of 20. For technical realisation, the platform uses leading (cloud) vendors such as OpenAI, Microsoft Azure, Google Cloud, AWS, XAI (Grok) and Mistral AI. All data is stored encrypted both during transmission and at rest, and privacy-by-design is leading throughout the platform. In view of using current Azure applications, the following standard security and network features are automatically in place:

• Ingress: automatic HTTPS/TLS protection, DDoS protection and load balancing, automatic certificate management, IP restrictions (allowlist/denylist), and integration capabilities with Web Application Firewall (WAF) for potentially malicious traffic.
• Egress: container apps run in sandboxed environment with limited outbound traffic; VNet integration possible via Virtual Network Integration (NSG/route-table control), including options for NAT gateway, Network Security Groups and Firewall-dependent outbound traffic.

To counter abuse and malicious traffic, a strict rate limiter has been set on all internal APIs of the chatbot. A maximum of 30 requests per minute can be performed per session. If a user or automated system exceeds this limit, further requests are automatically blocked for the remaining time period. This measure prevents platform overload and protects both the infrastructure and the service from DDoS attacks and other forms of unwanted traffic.

External provider

Chapter Data