GraphRAG

Are your agents still using old-fashioned RAG when they could be using GraphRAG?

In my opinion, LLMs have been over hyped, and vector indexes are where the real innovation will be seen.

But vector indexes alone have their limits.

That is where GraphRAG comes into play.

You take the creative search ability of Vector search and then ground it back to reality using graph relationships.

Let's say you wanted to search for chemical compounds that work well together(forgive me for this example; I am not a chemist).

You could enter in your query that does a first pass on the vector DB to get compounds that match your original query, but then what? You want to find compounds that don’t probabilistically work with the compound you found but actually have a definitive relationship with the compound you queried.

That is where the graph part comes into play.

Having the relationships between each record in our database mapped out with a graph DB like neo4j, which treats each record like a node and draws relationships between them, allows you to deterministically find related items.

This is a great balance between the probabilistic nature of vector search and the deterministic nature of graph search.

My question for you:

What technologies/techniques are you using for RAG?