TD-RAG: A Transparency-driven Retrieval-augmented Generation Framework with Multi-hop Task Planning and Selective Extraction

Chen Hao^1,2,3, Zhao Zhuofeng^1,2,3

1.North China University of Technology; 2.Beijing Key Laboratory on Key Technologies for AI+ Domain Applications; 3.Beijing Key Laboratory on Integration and Analysis of Large-scale Stream Data

Abstract：

Retrieval-Augmented Generation (RAG) addresses knowledge-intensive tasks by combining information retrieval with language generation. To overcome limitations in query semantics and reranking within current frameworks, this paper proposes Transparency-Driven RAG (TD-RAG). Unlike highly integrated platforms, TD-RAG offers a customizable architecture to suit complex applications. The framework introduces two core innovations: a multi-hop task planning model that decomposes complex queries into subtasks for precise retrieval, and an instruction-tuned selective extraction model for the reranking stage. This model identifies key information based on semantic relevance and intent rather than relying on rigid Top-k strategies. Experimental results demonstrate that TD-RAG significantly improves the alignment between retrieved content and actual requirements while enhancing system flexibility.

Key Words：

multi-hop search; Retrieval-Augmented Generation (RAG); re-ranking model; task planning