Les systèmes GraphRAG en pratique : vers une taxonomie des questions, des méthodes et défis d’évaluation

Retrieval-Augmented Generation (RAG) has emerged as an effective approach to enhance large language models by incorporating external knowledge, thereby improving their performance. However, traditional RAG, which primarily relies on textual corpora, exhibits several limitations, including the loss of global context due to document chunking and the lack of explicit modeling of entity relationships. In response to these limitations, GraphRAG aims to leverage graph structures to enhance the reasoning capabilities of LLMs. In this paper, we analyze existing GraphRAG methods, highlighting their specific characteristics. To complement this study, we propose a taxonomy of

questions and responses types, which are closely related to GraphRAG architectural choices. Finally, we provide an overview of evaluation methods for these systems, and more broadly for RAG systems.