Recent developments in the synergy between Large Language Models (LLMs) and Knowledge Graphs have introduced Graph Retrieval-Augmented Generation (Graph RAG) as an extension to traditional RAG. Graph RAG replaces the unstructured textual ‘knowledge base’ obtained in the retrieval phase of the RAG with Knowledge Graphs to improve entity identification and traceability. In this paper we present a customized Graph RAG framework for handling natural language queries over a knowledge graph with the help of Large Language Models. We conduct a comparative evaluation of several LLMs in responding to different query patterns on a domain-specific knowledge graph, employing four prompting levels defined by the TELeR taxonomy. The outputs are evaluated against ground truth results obtained by executing the corresponding SPARQL queries within a GraphDB triplestore. The content on which the experiments are performed is the Unified Medical Language System (UMLS) semantic network. The results summarize several strengths and weaknesses of the different LLMs relative to the complexity of selected query types, triple patterns, and prompting techniques involved in our Graph RAG setup.