Use Graph RAG when relationships are part of the evidence

When relationships are a crucial part of the evidence in answering a question, Graph RAG is an effective approach. It helps to establish the correct connections between facts and records that are essential for providing an accurate answer.

Vector retrieval is still highly valuable for finding similar documents or records based on the meaning of the question, even when the words used differ. It works well for a wide range of RAG applications when similarity alone can answer the query.

However, graph-based retrieval adds another layer of context by explicitly connecting related facts and records through typed, directed edges. Each edge is linked to its source and owner, along with confidence scores or effective dates, ensuring that the relationships are accurately represented. By combining graph RAG with vector retrieval, an agent can retrieve the necessary evidence and understand the connections between them, reducing the risk of errors or misinterpretations.

Graph RAG is commonly used when the question depends on relationships among facts, such as discovering which services rely on a particular library, who can approve an exception for an account, which customers are affected by a deployment, or which controls apply to specific data environments. By extracting relationships from the source data, graph RAG can provide a more complete and accurate answer to these types of questions.

It is important to note that while similar records may share semantic closeness, they might not necessarily have an operational connection. A graph RAG solution helps to clarify these relationships, ensuring that the answer is based on the correct connections and evidence. By leveraging both graph RAG and vector retrieval, systems can provide more reliable and comprehensive answers that accurately reflect the relationships and evidence present in the data.