In other words: you most likely had additionnal stuff in your variable conversation.summary.transcript which was passed to the knowledge agent and was used to generate an answer, hence the different answers depending on the different paths!