I found this supporting paper which helped define the process. I've implemented a simple Java component to generate the fragments as a SMILES string. I've used the Dynamic Java (on Server) so the Java code is embedded for you to look out (though I haven't really documented what I've done).
The component takes the incoming molecule and generates an array of SMILES strings for the child nodes. It uses the same SMARTS string that the paper uses to find the bonds to break. The order of child nodes is different because the SMILES generation is different - Daylight is used in the paper, whereas this obviously uses Pipeline Pilot. This means the order of the nodes found is different, but the actual nodes found are the same and match figure 1 in the original paper.
This code does not record parent relationships, nor does it add the other properties that the paper mentions, but I'm sure that those could be added.
I hope that's useful for you.