This paper proposes a novel post-hoc Citation-Enhanced Generation (CEG) approach for Large Language Models (LLMs) to reduce hallucinations in chatbots. Unlike previous methods that focus on preventing hallucinations during generation, CEG addresses the issue after generation by incorporating a retrieval module to search for supporting documents and a citation generation module to verify the factual accuracy of generated content. If a statement lacks supporting evidence, the model regenerates the response until all claims are supported by citations. CEG is a training-free, plug-and-play plugin that works with various LLMs. Experiments on three hallucination-related benchmarks show that CEG outperforms state-of-the-art methods in both hallucination detection and response regeneration. The framework includes a retrieval augmentation module, a citation generation module, and a response regeneration module. The retrieval module searches for relevant documents, the citation generation module uses natural language inference to verify the factual accuracy of claims, and the response regeneration module regenerates responses when necessary. The method is evaluated on multiple datasets, including WikiBio GPT-3, FELM, and HaluEval, demonstrating its effectiveness in reducing hallucinations. The results show that CEG achieves state-of-the-art performance in hallucination detection and regeneration. The framework is flexible and can be applied to various LLMs without additional training or annotations. The study also highlights the importance of citation in building responsible and accountable LLMs.This paper proposes a novel post-hoc Citation-Enhanced Generation (CEG) approach for Large Language Models (LLMs) to reduce hallucinations in chatbots. Unlike previous methods that focus on preventing hallucinations during generation, CEG addresses the issue after generation by incorporating a retrieval module to search for supporting documents and a citation generation module to verify the factual accuracy of generated content. If a statement lacks supporting evidence, the model regenerates the response until all claims are supported by citations. CEG is a training-free, plug-and-play plugin that works with various LLMs. Experiments on three hallucination-related benchmarks show that CEG outperforms state-of-the-art methods in both hallucination detection and response regeneration. The framework includes a retrieval augmentation module, a citation generation module, and a response regeneration module. The retrieval module searches for relevant documents, the citation generation module uses natural language inference to verify the factual accuracy of claims, and the response regeneration module regenerates responses when necessary. The method is evaluated on multiple datasets, including WikiBio GPT-3, FELM, and HaluEval, demonstrating its effectiveness in reducing hallucinations. The results show that CEG achieves state-of-the-art performance in hallucination detection and regeneration. The framework is flexible and can be applied to various LLMs without additional training or annotations. The study also highlights the importance of citation in building responsible and accountable LLMs.