TopoChat: Enhancing Topological Materials Retrieval with Large Language Models and Multi-Source Knowledge

Large language models (LLMs) perform well in general text tasks but face challenges in specialized fields like materials science. We present TopoChat, a knowledge-enhanced question-answering framework for materials science, which combines a domain-specific knowledge graph (TopoKG) and a literature clustering module. TopoChat retrieves both relevant subgraphs and literature information for each query, integrating structured and unstructured knowledge to support LLM reasoning. Experiments on two benchmarks, MaScQA and TopoQA, show that TopoChat improves answer accuracy across multiple LLMs. These results demonstrate that integrating knowledge graphs and literature context enhances reliability in scientific question answering. TopoChat provides an effective approach for adapting LLMs to complex domains, narrowing the gap between general language abilities and domain expertise.