Graph RAG vs SQL RAG

A performance comparison evaluated whether a SQL or a graph database works better for Retrieval-Augmented Generation (RAG) systems. Using a Formula 1 results dataset, various large language models were tasked with answering questions of increasing difficulty by querying both database types. The results showed that model capability was the most crucial factor, with the advanced GPT-5 achieving nearly perfect accuracy while older models struggled significantly. For the top-performing model, there was no meaningful difference in accuracy between using the structured SQL database and the relationship-focused graph database. This suggests that with modern LLMs, the choice of database can be guided by the data's natural structure rather than concerns about the model's query-generation ability.