HG-RAG Improves LLM Context Retrieval from Knowledge Graphs.

Traditional RAG systems often struggle when retrieving context from flat document stores for queries requiring complex hierarchical or relational reasoning. This new framework, HG-RAG, addresses this limitation by leveraging structured knowledge graphs. It navigates these graphs, expanding context through parent, neighbor, and child nodes based on a named entity anchor from the query. Evaluations show that HG-RAG consistently surpasses dense retrieval baselines across various scales and query types, particularly for hierarchical, relational, and multi-hop reasoning tasks. The system also demonstrates improved hallucination reduction and maintains locality coherence, offering a more robust approach to context retrieval for LLMs.