AutoThinkRAG: Advanced Control for Image-Text Interaction

Document Question Answering (DocQA) often struggles with lengthy contexts and excessive information, which complicates the reasoning process for Vision-Language Models (VLMs). While multimodal GraphRAG has made some progress, it still encounters significant obstacles: the need for large models to manage queries of varying complexities and the reasoning limitations of traditional VLMs. To tackle these challenges, we introduce AutoThinkRAG, a novel framework designed to improve the handling of complex documents through the collaboration of multiple models. A key feature is the Query Complexity Router, which directs reasoning pathways based on the complexity of the queries. Additionally, we propose a decoupled architecture where a compact VLM acts as a precise visual interpreter, converting relevant visual information into text. This text is then analyzed by a larger language model (LLM) for logical reasoning and synthesis. Our extensive testing on DocBench and MMLongBench shows that AutoThinkRAG not only lowers inference costs but also sets new benchmarks in performance. Further analysis confirms the effectiveness of our approach.