Efficacy of GraphRAG (Knowledge Graph-enhanced RAG) beyond Keyword Matching for Information Retrieval
Year: 2025
Type: dissertation
Abstract: Information retrieval systems face substantial difficulties because of the growing digital in-formation volume. The keyword-based methods face difficulties with natural language com-plexities and the vector-based Retrieval-Augmented Generation (RAG) system depends mainly on semantic similarity between individual text segments which restricts its effective-ness for queries that need to understand broad themes and interconnections and distributed knowledge in a corpus. This research explores Microsoft GraphRAG as a knowledge graph-enhanced RAG framework to address current retrieval system limitations. The GraphRAG system builds entity and relationship knowledge graphs through Large Language Models before creating hierarchical community summaries from unstructured text data. The devel-oped structures enable improved retrieval operations. The distinct retrieval strategies of Gra-phRAG specifically its Global and Local Search methods improve information retrieval for queries that need entity disambiguation and relationship-based retrieval and inferential capa-bilities and contextual understanding beyond standard vector-based RAG. The experiment used Microsoft GraphRAG and Traditional RAG baseline to process a medical bipolar dis-order knowledge base through a general query evaluation of retrieved context characteristics and quality. The preliminary analysis of a single query execution shows GraphRAG produc-es two distinct types of relevant context which include high-level synthesized points and de-tailed relation-rich entity structures instead of the raw text snippets from the baseline. The research demonstrates that knowledge graph implementation in GraphRAG creates new pos-sibilities to handle complex information retrieval operations which need reasoning capabili-ties above basic semantic matching thus improving RAG performance for sophisticated que-ries. The restricted evaluation scope requires additional extensive testing to establish defini-tive results about the hypotheses.
Author: Mai, Hoàng
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7000064150