In standard GraphRAG, a chatbot's answer to a question can only be given by using specific localities within the graph, which limits understanding of the wider context of the graph. Motivated by the Microsoft GraphRAG approach, we propose a solution for indexing and searching (partially built on the Memgraph-LlamaIndex extension) that can address this issue. By using hierarchical clustering over the knowledge graph (via the Leiden algorithm), we can handle complex queries that need big-picture answers, like identifying overarching themes in a dataset. It organizes data into meaningful clusters at different levels of detail, and then summarizes them to provide clear, context-aware chunks of information. Thus, when a user asks a question it is possible to give answers that show understanding over the full dataset at various levels of granularity.

Jacob Coles is a graduate Mechatronics Engineer from Australia who went on to complete a Masters of Language Technology at the University of Gothenburg in Sweden. He has then gone to work at Redfield AB where he has assisted in the development of APIs and applications which utilise machine learning and LLMs to build different types of RAG (Retrieval-Augmented Generation) applications that can fuse disparate data sources.

How would Microsoft GraphRAG work alongside a graph database?