The answer is that both are important. The trick is extracting the nodes and edges from the text automatically (which can be done with transfer learning from a foundation model) and then embed the string so you can search for things like it later, using the keyterms you find (and which are stored in the graph).