The term Graph RAG has become quite the buzzword in the industry lately, due to the popularity of knowledge graphs in “grounding” LLMs with domain-specific factual information.

The aim of this talk is to deconstruct Graph RAG into its components. A brief history of the literature that led to this term emerging is discussed, followed by some high-level architectures that represent how graphs can be used as part of RAG systems.

We summarize two examples of real-world projects that showcase tangible improvement in retrieval results due to their use of Graph RAG. Graph construction and obtaining high-quality graphs remain a bottleneck in building powerful Graph RAG systems, so some NLP techniques that assist these processes, such as named entity recognition, entity linking, and entity resolution are highlighted.