Paper Type
Short
Paper Number
PACIS2026-1559
Description
In knowledge-intensive question answering (QA) tasks, addressing semantically complex queries often requires integrating information from multiple documents and performing multi-step reasoning. Although Retrieval-Augmented Generation (RAG) method and its variants have shown promising performance in this domain, existing approaches often rely on rigid and static retrieval pipelines, limiting their ability to dynamically adjust reasoning strategies during execution. To overcome these limitations, we propose a feedback-aware and structurally stable Agentic RAG architecture that supports dynamic yet controlled revision of reasoning plans during task execution. The proposed architecture comprises five specialized agents—Planner, Tracker, Replanner, Dispatcher, and Executor—and represents task plans as Directed Acyclic Graphs (DAGs). This structured representation explicitly models reasoning dependencies and enables localized plan modifications without disrupting previously executed nodes. Experimental evaluations on two knowledge-intensive QA benchmarks demonstrate that the proposed approach improves both accuracy and robustness compared with strong baseline methods.
Recommended Citation
Hwang, San-Yih; Tsai, Ping-Yu; and Chang, Chia-Ming, "An Agentic RAG Architecture for Knowledge-Intensive QA" (2026). PACIS 2026 Proceedings. 10.
https://aisel.aisnet.org/pacis2026/ai_ml/ai_ml/10
An Agentic RAG Architecture for Knowledge-Intensive QA
In knowledge-intensive question answering (QA) tasks, addressing semantically complex queries often requires integrating information from multiple documents and performing multi-step reasoning. Although Retrieval-Augmented Generation (RAG) method and its variants have shown promising performance in this domain, existing approaches often rely on rigid and static retrieval pipelines, limiting their ability to dynamically adjust reasoning strategies during execution. To overcome these limitations, we propose a feedback-aware and structurally stable Agentic RAG architecture that supports dynamic yet controlled revision of reasoning plans during task execution. The proposed architecture comprises five specialized agents—Planner, Tracker, Replanner, Dispatcher, and Executor—and represents task plans as Directed Acyclic Graphs (DAGs). This structured representation explicitly models reasoning dependencies and enables localized plan modifications without disrupting previously executed nodes. Experimental evaluations on two knowledge-intensive QA benchmarks demonstrate that the proposed approach improves both accuracy and robustness compared with strong baseline methods.
Comments
01-AIML