DeepRead: 文档结构感知推理以增强智能体搜索

The rapid advancement of tool-using and agentic large language models (LLMs) is transforming Retrieval-Augmented Generation (RAG) from one-shot, passive retrieval into multi-turn, decision-driven evidence acquisition. While demonstrating strong performance in open-domain scenarios, current agentic search frameworks frequently treat lengthy documents as flat collections of chunks, thereby underutilizing inherent document priors such as hierarchical organization and sequential discourse structure. DeepRead introduces a structure-aware, multi-turn document reasoning agent specifically designed to operationalize these priors for long-document question answering. DeepRead employs an LLM-based OCR model to convert PDFs into structured Markdown, meticulously preserving headings and paragraph boundaries. This structured representation provides agents with richer contextual information, enabling more nuanced evidence retrieval and reasoning. By integrating both hierarchical and sequential document structures, DeepRead significantly enhances performance and accuracy in complex long-document question answering tasks. This innovative approach addresses the limitations of traditional RAG frameworks, which often suffer from information loss and insufficient context when processing extensive documents. DeepRead paves a new path for agents to achieve efficient and precise knowledge acquisition within intricate information environments, demonstrating a substantial leap in document understanding and reasoning capabilities for AI agents.