Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval

With the rise of generative AI, the way users access information has fundamentally changed. Search engines increasingly augment traditional results with AI-generated summaries, a paradigm exemplified by Google's AI Mode, which retrieves, reasons over, and synthesizes information from multiple web sources. Understanding and optimizing for this new retrieval paradigm is critical for content creators, marketers, and organizations that depend on search visibility.

Retrieval-Augmented Generation (RAG) has emerged as the dominant architecture for grounding large language model (LLM) outputs in factual, up-to-date information. However, most RAG implementations treat documents as unstructured text, discarding the rich structured metadata that many websites already provide via Schema.org markup and knowledge graph representations. This paper investigates whether structured linked data can improve RAG system performance and proposes a novel method to leverage this data.

The prevalent issue with current RAG systems is that they treat documents as flat text, ignoring the structured metadata and linked relationships provided by knowledge graphs. This approach leads to inadequate retrieval accuracy and completeness, especially in complex retrieval tasks requiring multi-source information integration. Addressing this problem is crucial for enhancing RAG system performance, particularly in applications requiring high precision and completeness.

The core innovation of this paper is the proposal of using structured linked data as a memory layer in RAG systems. Specifically, the paper effectively utilizes structured data from knowledge graphs through an enhanced entity page format, which includes llms.txt-style agent instructions, breadcrumbs, and neural search capabilities. This method significantly improves retrieval accuracy and answer quality compared to existing RAG systems, providing a new way of information integration.

The methodology of this paper includes the following key steps:

• �� Use Schema.org markup and dereferenceable entity pages as sources of structured linked data.
• �� Combine Vertex AI Vector Search 2.0 for retrieval with Google Agent Development Kit (ADK) for agentic reasoning.
• �� The experimental design tests seven conditions, covering three document representations (plain HTML, HTML with JSON-LD, and an enhanced agentic-optimized entity page) and two retrieval modes (standard RAG and agentic RAG with multi-hop link traversal), plus an Enhanced+ condition.
• �� Experiments are conducted across four domains (editorial, legal, travel, e-commerce) to validate the method's effectiveness.

The experimental design covers four domains (editorial, legal, travel, e-commerce), using Vertex AI Vector Search 2.0 for retrieval and combining with Google Agent Development Kit (ADK) for agentic reasoning. The experiments include seven conditions, covering three document representations and two retrieval modes, plus an Enhanced+ condition. The dataset includes 2,443 individual query evaluations, ensuring the reliability and reproducibility of the results.

The experimental results show that the enhanced entity page format significantly improves retrieval accuracy: +29.6% for standard RAG and +29.8% for the full agentic pipeline. The Enhanced+ variant achieves the highest absolute scores in accuracy and completeness (accuracy: 4.85/5, completeness: 4.55/5), though the incremental gain over the base enhanced format is not statistically significant. This demonstrates the advantage of structured data in multi-hop link traversal.

This method can be directly applied to high-precision and high-completeness applications such as legal document retrieval, travel information integration, and e-commerce product recommendation. By leveraging structured linked data, the method can significantly enhance retrieval accuracy and completeness, especially in complex retrieval tasks requiring multi-source information integration.

Despite the excellent performance of this method in experiments, there are some limitations. First, although the Enhanced+ variant achieves the best absolute scores, the incremental gain over the base enhanced format is not statistically significant, indicating that further navigational affordances may not always lead to significant performance improvements. Additionally, the experimental results may be influenced by specific domain datasets, with more pronounced effects in domains rich in knowledge graph information. Finally, due to the complexity of the experimental design, the reproducibility of results may be limited, especially across different knowledge graphs and data platforms.