With the growth and popularity of large-scale language models (LLMs), chatbots have become essential to transforming interactions between business users through specialization, personalization, efficiency, and scalability. Nevertheless, issues such as hallucinations, specialization limits, and maintaining accuracy while multitasking remain critical challenges. In response to these concerns, this paper suggests a modular architecture for smart chatbots with LLMs as dynamic routers to route user queries to expert agents. The idea is to reduce hallucinations while increasing the relevance and accuracy of responses and show the strength of more specialist prompts. The suggested architecture classifies agents by domain (health and finance) and then by subdomains like symptoms or money planning. More advanced prompting techniques, including chain-of-thought and few-shot learning, are employed to improve contextual understanding and response accuracy. Furthermore, the findings indicate that using different language models for different situations can maximize query routing and maximize the level of detail in responses, which will be better matched to user needs. Experimental results confirm that the modularity-based solution is viable under multi-domain conditions and reduces hallucinations significantly while enhancing outputs accuracy and contextualization.