Graf von Data: A Knowledge Graph Question Answering Agent for Organisational Usage

Large Language Model-based (LLM) systems have advanced Knowledge Graph-based Question Answering (KGQA) but overlook organisational demands for transparency, control, and flexible deployment. We present Graf von Data (GvD), an LLM-augmented agent for KGQA, designed to meet key organisational demands. GvD translates natural language questions into SPARQL queries by iteratively exploring the knowledge graph (KG) using the ReAct framework and a set of three actions — search, describe, and query — requiring only a SPARQL endpoint with text search capabilities. Unlike other approaches, GvD is graph-agnostic and uses an open-weight LLM. We evaluate GvD in the Text2SPARQL challenge, demonstrating near state-of-the-art performance on the Corporate KG benchmark. In addition to quantitative results, we conduct a qualitative comparison against the challenge competitors, showing that GvD thrives under the organisational demands.