They seem to be using RAGs to prevent hallucinations (which I imagine would be really bad in their context).
The 13B Llama 2 model is not that good compared to the best ones. Maybe it was easier for them to fine tune?
They do mention that their expectation is that the 70B model will provide even better performance. I expect that you are correct and that they determined the 13B was capable enough to serve as a base model. Why incur additional training time before getting preliminary results.
This reminds me of a Star Trek (The Next Generation) character consulting an AI expert on the Holodeck.
It's funny how the sttng model of computer-assisted research has gone from “laughably naive about what a computer even is” to “available to anyone for 20 dollars a month” in less than two years (from the consumer perspective).