# how use RBF for primal model of svm?

I know if we want to solve primal model of non-linear SVM, we have to generate new features. for example for kernel (1+xz)^2 for primal problem for any pair of features x1 and x2 we have to generate:

   (x1,x2) -> (x1,x2,x1^2, x2^2, sqrt(2)*x1*x2)


I just don't know how are new features when we use RBF. I know the kernel for lagragian is

   gamma* exp((x1-x2)^2)


how about if we want to solve primal problem? I mean what are new features?