We propose DynamicRAG, which optimizes retrieval-augmented generation by using reinforcement learning to adaptively rerank documents based on LLM's responses as feedback, achieving state-of-the-art results across multiple datasets.