This paper proposes a retrieval augmented response generation (RARG) framework for generating counter-misinformation responses. Online misinformation poses significant threats to public interest, and existing methods often produce low-quality, repetitive responses. RARG addresses this by collecting scientific evidence from a large database of academic articles and generating responses based on this evidence. The framework consists of two stages: (1) evidence collection, where a retrieval pipeline is used to find and rank relevant evidence documents; and (2) response generation, where large language models (LLMs) are aligned with reinforcement learning from human feedback (RLHF) to generate evidence-based, polite, and factual responses. The RARG framework is evaluated on both in-domain and cross-domain datasets, demonstrating superior performance compared to existing methods. The results show that RARG consistently generates high-quality counter-misinformation responses, particularly in the context of COVID-19 misinformation. The framework's two-stage retrieval and fine-grained RLHF alignment enable effective evidence-based response generation, making it a promising solution for combating online misinformation.This paper proposes a retrieval augmented response generation (RARG) framework for generating counter-misinformation responses. Online misinformation poses significant threats to public interest, and existing methods often produce low-quality, repetitive responses. RARG addresses this by collecting scientific evidence from a large database of academic articles and generating responses based on this evidence. The framework consists of two stages: (1) evidence collection, where a retrieval pipeline is used to find and rank relevant evidence documents; and (2) response generation, where large language models (LLMs) are aligned with reinforcement learning from human feedback (RLHF) to generate evidence-based, polite, and factual responses. The RARG framework is evaluated on both in-domain and cross-domain datasets, demonstrating superior performance compared to existing methods. The results show that RARG consistently generates high-quality counter-misinformation responses, particularly in the context of COVID-19 misinformation. The framework's two-stage retrieval and fine-grained RLHF alignment enable effective evidence-based response generation, making it a promising solution for combating online misinformation.