Abstract
Large language models (LLMs) have shown remarkable multimodal informationprocessing and reasoning ability. When equipped with tools through functioncalling and enhanced with retrieval-augmented techniques, compound LLM-basedsystems can access closed data sources and answer questions about them.However, they still struggle to process and reason over large-scalegraph-structure data. We introduce the GDS (Graph Data Science) agent in thistechnical report. The GDS agent introduces a comprehensive set of graphalgorithms as tools, together with preprocessing (retrieval) and postprocessingof algorithm results, in a model context protocol (MCP) server. The server canbe used with any modern LLM out-of-the-box. GDS agent allows users to ask anyquestion that implicitly and intrinsically requires graph algorithmic reasoningabout their data, and quickly obtain accurate and grounded answers. Weintroduce new benchmarks that evaluate intermediate tool calls as well as finalresponses. The results indicate that GDS agent is able to solve a wide spectrumof graph tasks. We also provide detailed case studies for more open-ended tasksand study scenarios where the agent struggles. Finally, we discuss theremaining challenges and the future roadmap.