Have a close look at the csaps documentation. There are two kinds of weighting factor you can see in the equation under "this smoothing spline f minimizes:", the first, w, operates directly on the squared error, the other, (1-p)*lambda operates on the 2nd derivative of f.
If you are just looking to weight the data with error bars (if y2 are your 1-sigma error bars), for example, you should just input:
pp = csaps(x,y,1,x,1./y2.^2);
If instead you really do want y2 to weight the smoothness of different areas (i.e set p not equal to 1), then you need to create a vector with a p value as the first element, and the 2:end elements as the lambda weighting factor for each knot of x, for example:
pp = csaps(x,y,[0.99,y2(2:end)]);
Again, look closely at the documentation, it makes sense when you sit down and read it all the way through.
