Domen Hočevar and Klemen Kenda

Abstract
Knowledge graphs have traditionally required the use of specific
query languages, such as SPARQL, to retrieve relevant data. In
this paper, we present a system capable of performing natural
language queries on knowledge graphs by leveraging retrieval-
augmented generation (RAG) and large language models (LLMs).
Our system can ingest large knowledge graphs and answer queries
using two approaches: first, by utilizing LLMs to extract informa-
tion directly from subgraphs; and second, by generating SPARQL
queries with LLMs and using the results to inform further infer-
ence, such as counting the number of items.