Analyzing the Efficacy of an LLM-Only Approach for Image-based Document Question Answering