Recent advancements in large language models (LLMs) have greatly enhanced natural language processing capabilities, particularly in open-domain question answering and dialogue generation. However, LLMs still face challenges in knowledge-intensive tasks, such as generating accurate content. Traditional retrieval-augmented generation (RAG) models address some of these issues by retrieving external knowledge. However, standard RAG often retrieves incomplete or overly redundant information, harming generation quality. To overcome these limitations, we propose C2-RAG (Context Completeness RAG), a framework designed to ensure the completeness of context information. The innovation of C2-RAG lies in its ability to dynamically assess whether the retrieved context sufficiently answers the user’s query. When key information is missing, C2-RAG initiates additional retrieval to supplement the gaps. By utilizing a context extractor and a context sufficiency evaluator, C2-RAG iteratively refines the context provided to the model, reducing the impact of irrelevant or redundant information and ultimately improving the quality of the generated answers. Our approach is highly adaptable across different tasks, requiring no fine-tuning of the generation model. Experimental results show that C2-RAG significantly improves answer accuracy across several knowledge-intensive question-answering benchmarks, demonstrating its effectiveness and potential for practical applications.