Enhancing Pipeline-Based Conversational Agents with Large Language Models