This paper discusses Retrieval Augmented Generation (RAG), which has become a fundamental paradigm for addressing the challenges faced by large-scale language models (LLMs) in processing real-time information and domain-specific problems. Existing RAG systems primarily rely on the in-context learning (ICL) capabilities of the LLM itself, but in-depth research on the specific capabilities required for RAG generation models is lacking, leading to inconsistent document quality and flawed retrieval systems. Even the limited research on fine-tuning RAG generation models lacks a granular focus on RAG tasks or a deep understanding of the thought chain process. To address this, this paper proposes that RAG models should possess three progressively hierarchical capabilities: (1) filtering: the ability to select relevant information; (2) combining: the ability to combine semantic information across paragraphs; and (3) RAG-specific inference: the ability to further process external knowledge using internal knowledge. Therefore, we present Hierarchical Thinking-Directed Adjusted Retrieval Augmented Generation (HIRAG), a novel RAG-directed fine-tuning method that incorporates a "think before answer" strategy. This method leverages a multi-stage progressive thought chain to enhance the model's open-book testability. Experimental results show that the HIRAG training strategy significantly improves model performance on datasets such as RGB, PopQA, MuSiQue, HotpotQA, and PubmedQA.
