Retrieval-augmented Multi-modal Chain-of-Thoughts Reasoning for Large Language Models