Structured Knowledge Representation through Contextual Pages for Retrieval-Augmented Generation

TL;DR

This work tackles the challenge of organizing external knowledge within Retrieval-Augmented Generation by introducing PAGER, a page-driven autonomous knowledge representation framework. PAGER prompts an LLM to generate a cognitive outline with knowledge slots and then iteratively fills these slots with retrieved evidence to form a structured contextual page used to guide QA. Across multiple knowledge-intensive benchmarks and backbone models, PAGER consistently outperforms baselines, demonstrating higher-quality, information-dense representations, reduced knowledge conflicts, and improved knowledge utilization by LLMs. The approach offers a cognitively grounded, scalable method to unify retrieval and reasoning, with practical implications for robust multi-hop QA and other knowledge-intensive tasks.

Abstract

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge. Recently, some works have incorporated iterative knowledge accumulation processes into RAG models to progressively accumulate and refine query-related knowledge, thereby constructing more comprehensive knowledge representations. However, these iterative processes often lack a coherent organizational structure, which limits the construction of more comprehensive and cohesive knowledge representations. To address this, we propose PAGER, a page-driven autonomous knowledge representation framework for RAG. PAGER first prompts an LLM to construct a structured cognitive outline for a given question, which consists of multiple slots representing a distinct knowledge aspect. Then, PAGER iteratively retrieves and refines relevant documents to populate each slot, ultimately constructing a coherent page that serves as contextual input for guiding answer generation. Experiments on multiple knowledge-intensive benchmarks and backbone models show that PAGER consistently outperforms all RAG baselines. Further analyses demonstrate that PAGER constructs higher-quality and information-dense knowledge representations, better mitigates knowledge conflicts, and enables LLMs to leverage external knowledge more effectively. All code is available at https://github.com/OpenBMB/PAGER.

Figures (13)

• Figure 1: The Pipeline of Our PAGER Model. PAGER initializes a blank structured page with multiple slots and preliminarily fills the retrieved external knowledge into the corresponding slots, thereby constructing more comprehensive and coherent knowledge representations.
• Figure 2: The Illustration of Our PAGER Model.
• Figure 3: The Quality of Knowledge Representations Constructed by Different Methods.
• Figure 4: Slot Ablation Studies on the Constructed Page. "N/A" denotes the complete page with no filled slots removed. "First", "Second", "Third", and "Fourth" denote the variants in which the First, Second, Third, and Fourth filled slots are removed, respectively. PAGER (w/o Outline) denotes a variant where the cognitive outline structure is removed from the complete page and is used as a baseline for comparison.
• Figure 5: Effects of Different Knowledge Representations on Retrieval and Generation Modules of RAG.