Skip to content
RAGnaroXSecure on-premises AI
DE EN

Technology & Research

Scientifically evaluated, not just promised

RAGnaroX was developed together with the University of Innsbruck and published as a peer-reviewed paper at ICSE 2026, one of the leading international conferences for software engineering.

Benchmark results at a glance

The evaluation used established datasets that reflect typical practical challenges: SQuAD v1.1 for simple factual questions, MultiHopRAG for multi-step questions across documents and MLQA for cross-language questions. Measurements used the RAGAS framework on an NVIDIA RTX 4090.

The compact Qwen3 model with 4 billion parameters achieved the highest faithfulness and answer relevancy for simple questions. More model size does not automatically mean better answers from your data. For multi-hop questions, retrieval quality is comparable to multilingual embedding models; knowledge-graph integration is the next development step.

Practical test with real customer data

Beyond standard benchmarks, RAGnaroX was tested on a documentation-based use case from CASABLANCA hotelsoftware with 250 question-answer pairs. The cloud-based production system achieved 7% higher context recall@10, while answer relevancy was nearly identical. RAGnaroX can process real company data without cloud services or commercial model APIs.

GENIUS research project

RAGnaroX is being developed as part of the ITEA4 GENIUS research project, funded by FFG under grant number 921318, in cooperation with the Department of Computer Science at the University of Innsbruck. The paper appears in the proceedings of the 48th International Conference on Software Engineering.