What is SPARK API?

In the SPARK project ("Schnellere Planung und Realisierung durch KI" – Faster Planning and Implementation through AI), the Federal Ministry for Digital Transformation and Government Modernisation commissioned the development of the AI solution SPARK as part of an innovation partnership. The development was financed by the Climate and Transformation Fund of the Federal Ministry for Economic Affairs and Energy.
SPARK API provides combinable, domain-adaptable AI services for document processing, formal review, objection management, material legal assessment, and AI-assisted decision drafting. The foundation is a versioned, interconnected legal knowledge graph that structures norms, evidence references, and decision options, continuously updated via interfaces to legal sources such as NeuRIS and GII. Explainability, verifiability, and human-in-the-loop mechanisms ensure traceable decisions, review steps, and norm references — with no black-box character. Full modularity is achieved through Kubernetes, microservices, micro-frontends, and an API-first paradigm.

Architecture & Platform

SPARK API exposes workflow-agnostic domain services via APIs that can be used independently of any overarching orchestration, making them suitable for flexible integration into existing platforms and specialised administrative systems. The solution is platform-independent (cloud / on-premise) and follows a three-layer model of service, integration, and AI layers. An API gateway handles routing, authentication, and policy enforcement; asynchronous queue-based processing (RabbitMQ) decouples load spikes.

Integration & Deployment Model

Integration is loosely coupled via standardised APIs (REST, messaging, XÖV). SPARK API can be combined with existing systems, processes, and UIs, and supplemented by external orchestration (e.g. specialised administrative systems or workflow engines). Micro-frontends enable modular embedding of individual UI components; consuming the bundled UI is not required.

AI Architecture & Data Processing

The focus is on production-ready AI services without in-house model training. Embedding, reranker, and language models are designed for straightforward replacement. A knowledge graph models legal knowledge hierarchically and combines it with a RAG architecture. AI-assisted functions include coherence and completeness checks, handling of objections, legal fact assessment, and decision drafting. Transparency is ensured through explainability, auditability, human-in-the-loop mechanisms, and statistical analyses. Models can be operated in-house or consumed as a model-as-a-service.

Operations & Security

Deployment on Kubernetes with a full DevOps infrastructure (Helm, ArgoCD, GitOps); horizontal scaling via containers and services; tenant isolation via separate deployments. Security is provided by OAuth2 / OIDC (Keycloak), mTLS-secured service communication (Istio), rate limiting at the API gateway, and centralised secrets management.

Project Status and Open Source Release

SPARK API has been published as open source. The solution is being tested with a first federal authority and is being continuously improved.
A detailed overview of the functions can be found here.