AI4DeliberationD4. 1 – AI toolkit, Architecture and Knowledge Graph – first version

This deliverable (D4.1) reports on the technical infrastructure of the AI4Deliberation project, consolidating the work of the first two sprints of Work Package 4 (Tasks T4.1 and T4.2). It documents the technical foundations that enable AI4Deliberation's AI-assisted, multimodal, and gamified deliberations. In line with the Grant Agreement, the deliverable presents three main contributions: · WP4 AI toolkit and sandbox used to identify, evaluate, and operationalise open-source AI tools and models (including LLMs and multimodal/video-capable models) as deliberation services. The sandbox provides a controlled and repeatable testing environment across pilots and languages, supporting evidence-driven refinement and exposing AI functionality as reusable services via APIs. · First operational design of the Deliberation Knowledge Graph (DKG), including a dual-layer semantic model (ontological blueprint and FAIR data governance) and adoption of the IBIS framework as a core logical grammar to represent issues, positions, and arguments. The DKG layer is complemented by a Multi-Agent RAG ("Cognitive Retrieval") mechanism that translates natural-language questions into validated SPARQL queries. · First and refined versions of the Reference Architecture, evolving from the Technical Annex vision (v1) to an implementation-facing blueprint (v2) organised into distinct layers (Interaction, Data Preprocessing, Intelligence, and an enveloping Compliance layer), with explicit platform integration targets, formalised preprocessing readiness gates, and end-to-end compliance aligned with GDPR, AI Act considerations, and FAIR principles. WP4's implementation follows an agile, ethics-by-design methodology, applying the ALTAI framework as a shared operational reference for the EU's Trustworthy AI requirements. The deliverable also outlines ongoing T4.2 work to refine and (where needed) fine-tune and re-evaluate models, provide concrete AI services for pilots (e.g., deliberation summarisation and gamification), and deploy and populate a FAIR Deliberation Knowledge Graph hosted in an open-source graph database. The AI services covered include summarisation, argument extraction, moderation, translation, topic modelling, gamification, and conversational analysis, integrated across the four project pilots (German, Greek, International, and Italian).