The paper introduces a novel concept-based Retrieval-Augmented Generation (RAG) framework that leverages Abstract Meaning Representation (AMR) to enhance the performance of Large Language Models (LLMs) in handling long-tail, domain-specific queries. The proposed framework addresses the limitations of LLMs, which often rely on potentially flawed parametric knowledge, leading to hallucinations and inaccuracies. By incorporating external, non-parametric knowledge through RAG, the framework aims to improve the reliability and accuracy of LLMs' responses. The core contribution of the paper is the AMR-based concept distillation algorithm, which compresses long-context retrieved documents into a compact set of crucial concepts. This algorithm formats the informative nodes of AMR graphs, ensuring that LLMs focus on essential information during inference. The method is evaluated on open-domain question-answering datasets (PopQA and EntityQuestions), demonstrating superior performance compared to baseline methods, especially as the number of supporting documents increases. The results show that the proposed framework effectively reduces interference from irrelevant information and enhances the accuracy and robustness of LLMs in long-context scenarios. The paper also discusses related works on long-context understanding and linguistic-augmented Natural Language Generation (NLG), highlighting the advantages of AMR in capturing essential concepts. The experimental results and analysis provide insights into the effectiveness of the proposed method, showing its potential to improve RAG performance in various contexts. The conclusion emphasizes the novel direction offered by integrating structured semantic representations with RAG to handle tasks requiring high fidelity to the knowledge.The paper introduces a novel concept-based Retrieval-Augmented Generation (RAG) framework that leverages Abstract Meaning Representation (AMR) to enhance the performance of Large Language Models (LLMs) in handling long-tail, domain-specific queries. The proposed framework addresses the limitations of LLMs, which often rely on potentially flawed parametric knowledge, leading to hallucinations and inaccuracies. By incorporating external, non-parametric knowledge through RAG, the framework aims to improve the reliability and accuracy of LLMs' responses. The core contribution of the paper is the AMR-based concept distillation algorithm, which compresses long-context retrieved documents into a compact set of crucial concepts. This algorithm formats the informative nodes of AMR graphs, ensuring that LLMs focus on essential information during inference. The method is evaluated on open-domain question-answering datasets (PopQA and EntityQuestions), demonstrating superior performance compared to baseline methods, especially as the number of supporting documents increases. The results show that the proposed framework effectively reduces interference from irrelevant information and enhances the accuracy and robustness of LLMs in long-context scenarios. The paper also discusses related works on long-context understanding and linguistic-augmented Natural Language Generation (NLG), highlighting the advantages of AMR in capturing essential concepts. The experimental results and analysis provide insights into the effectiveness of the proposed method, showing its potential to improve RAG performance in various contexts. The conclusion emphasizes the novel direction offered by integrating structured semantic representations with RAG to handle tasks requiring high fidelity to the knowledge.