What is the difference between yfit and predict functions?

I have a Regression app generated GPR compact model saved on my desktop. During training I had PCA feature selection on. When I use the 'yfit' function with the model on out-of-sample data set (i.e., new data not included in the training stage) it is able to provide reasonable answers. So "yfit" uses the PCA companents embedded in the model structure. But when I use "predict" function it seems that it requires some preprocessing (i.e., projection of raw variables onto PCA components).

Why is that? What is the difference between the two?