How real is graph RAG, I see we are gradually evolving to GRAPH RAG

Community
1

Of late I am getting lot of use cases where it is not just the semantics, but the relationships and metadata associated with the data also becomes important. for the same question, the answer may be different based on the relationships between the question topics , chunks and user role. I am therefore seeing an importance of GRAPH RAG in the generative AI space. Traditional RAG, I feel is limiting. I wanted to take inputs and opinion from the community also. How are we seeing this space evolving?

1 Like
2

RAG strategy needs to be tailored to each use case, I do not believe there is a “best” approach and there will be trade offs.

This means choosing the right embedding models, chunking and retrieval method be that graph, bm25, vector or hybrids. Semantic search is still powerful for many use cases, knowledge graphs are expensive to build and maintain at scale.

The next wave will see your knowledge be part of the model via stacking (see DoRA [2402.09353] DoRA: Weight-Decomposed Low-Rank Adaptation). Use search retrieval only where needing transactional or live information.

1 Like