How does GraphRAG differ from vector RAG?

GraphRAG and vector RAG are approaches to Retrieval-Augmented Generation (RAG), which differ in the way they organize and search for information to generate responses.

• Vector RAG uses vector representations (embeddings) of documents or text fragments. When a query is made, it searches for the nearest neighbors in the vector space, and the relevant fragments found are passed to the generative model to create a response. This approach works well with unstructured data and scales effectively.

• GraphRAG extends the idea by using a knowledge graph, where information is represented as nodes and connections between them. The search for relevant data considers the graph's structure, allowing for the consideration of semantic and logical relationships between elements. This improves the quality of search and response generation, especially in complex domains with rich knowledge structures.

Thus, the main difference is that vector RAG operates with flat vector representations, while GraphRAG uses structured graph data, providing additional capabilities for contextualization and accuracy.