Reasonably than storing knowledge in rows and columns for conventional searches, or as embeddings for vector search, a information graph represents knowledge factors as nodes and edges. A node might be a definite truth or attribute, and edges will join all of the nodes which have related relationships to that truth. Within the instance of a product catalog, the nodes will be the particular person merchandise whereas the perimeters might be comparable traits that every of these merchandise possess, like measurement or coloration.
Sending a question to a information graph entails in search of all of the related entities to that search, after which making a information sub-graph that brings all these entities collectively. This retrieves the related info for the question, which might then be returned again to the LLM and used to construct the response. This implies which you could cope with the issue of getting a number of comparable knowledge sources. Reasonably than treating every of those sources as distinct and retrieving the identical knowledge a number of occasions, the information might be retrieved as soon as.
Utilizing a information graph with RAG
To make use of a information graph along with your RAG utility, you may both use an present information graph with knowledge that’s examined and identified to be appropriate prematurely, or create your personal. When you’re utilizing your personal knowledge—comparable to your product catalog—you’ll want to curate the information and test that it’s correct.