GraphRAG extracting themes from news articles

[rbrugaro]

In the GraphRAG youtube demo: https://youtu.be/r09tJfON6kE?t=547 it is shown that with GraphRAG approach performs more LLM calls (10 vs 1), tokens (49k vs 5k), and time (71s vs 8s) compared to the baseline RAG.
what determines the number of LLM calls? those calls include graph creation or those are iterations determined by agent?
for quality comparisons have you considered constraining the number of LLM calls to the same in baseline RAG and graphRAG? it would be good for understanding the effect of using knowledge graph vs. repeated LLM refinement steps in the final answer quality.

Thanks!