Reducing Semantic Noise in Historical RAG through Knowledge-Graph Constraints

This paper presents a comparative study of retrieval architectures applied to historical factual knowledge. We compare a purely semantic vector-based retrieval approach (SVR) with a knowledge-graph constrained retrieval method (KGC), integrating relational constraints from Wikidata and spatiotemporal filtering. The results show that ontology-guided retrieval significantly reduces semantic noise and improves the precision of evidence retrieval in historical queries. The study highlights the importance of structural constraints in retrieval systems for Digital Humanities and historical research contexts.