VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents