Document Question Answering (DocQA) aims to answer textual questions based on the information contained within the document. Effective integration of textual and visual cue features remains a challenge for current retrieval augmented generation (RAG) approaches. To address this issue, we propose M2RAG, a novel multimodal RAG framework. The framework integrates dual-tower retrieval and multi-agent generation mechanism to effectively deal with multi-modal retrieval, and uses text filter, visual extractor and modal fuser to reason collaboratively to achieve a comprehensive understanding of document content. To enhance the performance of the model in DocQA, we adopt a knowledge distillation based fine-tuning strategy to train a lightweight visual language model on the constructed fine-tuning dataset as the core component of the multi-agent module. Experiments on DocBench, MMLongBench, and LongDocURL show that M2RAG outperforms baseline methods by an average of 7.7%, demonstrating its effectiveness in complex document question answering.