we added mechanical layers where ever possible , Think of it as you wanted a piece of quality work who will you choose A single person or 4-5 diff people who cross check each other work

Thats the architecture we follow here but its not just we just took multiple model because in order to verify answers we had to implement some layers, Like for a code query it goes to diff layer, for reasoning , for summary, for factual and so on We tried to build this as mechanical as we can because a single Llm can't see its own mistakes unless someone points that out .