Knowledge-Graph-Guided Retrieval-Augmented Generation for Domain-Specific Question Answering

Authors

  • Lei Zan Author
  • Gang Yao Author
  • Meimei Zhang Author

Abstract

Retrieval-augmented generation (RAG) grounds large language models in an external corpus, but in specialised domains such as technical support and clinical enquiry the retrieved passages are fragmentary, redundant and frequently centred on surface wording rather than on the entities the question actually asks about. This paper presents KGRG, a knowledge-graph-guided retrieval-augmented generation framework that couples a domain knowledge graph with a generative reader. The framework first links question entities to the graph, retrieves a compact K-hop sub-graph whose candidate paths are scored by a learned relation-aware ranking function, and then fuses graph evidence with dense passage retrieval through a weighted scoring model. During generation, a constrained decoding step keeps entity mentions consistent with the retrieved sub-graph and attaches a citation to every factual span. Experiments on two domain benchmarks, TechQA and PubMedQA-derived clinical questions, show that KGRG improves exact-match scores by 8.7-14.7 points over a strong RAG baseline and by 3.1-7.4 points over a graph-oriented GraphRAG pipeline, while reducing the average length of the generated context by 38 percent. An ablation confirms that the gains stem mainly from relation-aware path scoring and constrained decoding. The results indicate that explicit structural evidence is a practical complement to dense retrieval when questions target specific domain entities.

Downloads

Published

2026-03-25

Issue

Section

Articles