Specialized and Auditable Knowledge Graphs for Enterprise AI

SAGE-AI tackles a core barrier to enterprise AI adoption in regulated industries: large language models (LLMs) cannot reliably extract verifiable and auditable knowledge from technical document corpora, while building knowledge graphs manually is slow and expensive. The experiment uses EuroHPC supercomputing resources to identify, train, and optimise configurations of multiple Small Language Models (SLMs) that automatically extract structured, multi-level knowledge from biomedical literature and assemble it into large, claim-centric knowledge graphs with full provenance and built-in auditability.

Coordinated by Deepentix (Austria) in collaboration with research partner TIB (Germany), the project combines multi-stage SLM pipelines with HPC-scale training. Its goal is to demonstrate that trustworthy, domain-specific AI can be enabled through hyperspecialisation at a fraction of the cost of large, general-purpose LLMs, while ensuring compliance with GDPR and the EU AI Act.

Organisations involved: 

End-User: Deepentix FlexCo

Technology expert: Leibniz Information Centre for Science and Technology - TIB