The team shipped some pretty incredible changes to how Ada’s knowledge retrieval and grounded generation works. I’m bound to secrecy on the technique, but what I can say is that it uses a novel process to index a derivative of articles alongside the articles themselves. This derivative improves first contact resolution by up to 8 percentage points in our testing. It completely rolled out today after a lengthy beta. I’m very proud of the collaboration between ML research and production application teams on this project and I expect it to push us one step closer to beating human performance at customer support.