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

Andrea Volpini; Elie Raad; Beatrice Gamba; David Riccitelli

cs.IRMar 11, 2026

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

Authors: Andrea Volpini, Elie Raad, Beatrice Gamba, David Riccitelli

Organizations: WordLift, Rome, Italy

Abstract

Retrieval-Augmented Generation (RAG) systems typically treat documents as flat text, ignoring the structured metadata and linked relationships that knowledge graphs provide. In this paper, we investigate whether structured linked data, specifically Schema.org markup and dereferenceable entity pages served by a Linked Data Platform, can improve retrieval accuracy and answer quality in both standard and agentic RAG systems. We conduct a controlled experiment across four domains (editorial, legal, travel, e-commerce) using Vertex AI Vector Search 2.0 for retrieval and the Google Agent Development Kit (ADK) for agentic reasoning. Our experimental design tests seven conditions: three document representations (plain HTML, HTML with JSON-LD, and an enhanced agentic-optimized entity page) crossed with two retrieval modes (standard RAG and agentic RAG with multi-hop link traversal), plus an Enhanced+ condition that adds rich navigational affordances and entity interlinking. Our results reveal that while JSON-LD markup alone provides only modest improvements, our enhanced entity page format, incorporating llms.txt-style agent instructions, breadcrumbs, and neural search capabilities, achieves substantial gains: +29.6% accuracy improvement for standard RAG and +29.8% for the full agentic pipeline. The Enhanced+ variant, with richer navigational affordances, achieves the highest absolute scores (accuracy: 4.85/5, completeness: 4.55/5), though the incremental gain over the base enhanced format is not statistically significant. We release our dataset, evaluation framework, and enhanced entity page templates to support reproducibility.

Figures & tables

ID	Document Format	Retrieval Mode	Hypotheses
C1	Plain HTML	Standard RAG	H1 baseline
C2	HTML + JSON-LD	Standard RAG	H1 treatment
C3	Enhanced entity	Standard RAG	H3 baseline
C4	Plain HTML	Agentic RAG	H2 baseline
C5	HTML + JSON-LD	Agentic RAG	H2 treatment
C6	Enhanced entity	Agentic RAG	H2+H3 treatment

Table 1: Experimental conditions.

Figure 1: Before and after: plain HTML (left) vs. enhanced entity page (right) for a sample entity. The enhanced format adds structured breadcrumbs, related entity links with dereferenceable URIs, agent instructions in llms.txt style, and an embedded JSON-LD block—yielding a +29.6% accuracy improvement in standard RAG and +29.8% in the agentic pipeline.

Figure 2: System architecture. User queries are processed by a Google ADK agent that orchestrates three tools: vector search over Vertex AI, entity link traversal, and neural search via MCP. Documents are indexed in three formats (C1–C6) and the agent generates grounded answers using a ReAct-style reasoning loop.

ID	Condition	Accuracy	Completeness
C1	Plain HTML, Std.	3.62 $\pm$ 1.82	3.01 $\pm$ 1.94
C2	HTML+JSON-LD, Std.	3.89 $\pm$ 1.70	3.33 $\pm$ 1.85
C3	Enhanced, Std.	4.69 $\pm$ 0.95	4.45 $\pm$ 1.25
C4	Plain HTML, Agent.	4.36 $\pm$ 1.33	3.98 $\pm$ 1.60
C5	HTML+JSON-LD, Agent.	4.40 $\pm$ 1.21	4.00 $\pm$ 1.54
C6	Enhanced, Agent.	4.70 $\pm$ 0.82	4.38 $\pm$ 1.20

Table 2: Main results across experimental conditions (mean

\pm

std).

Figure 3: Mean accuracy and completeness scores by experimental condition. Enhanced entity pages (C3, C6, C6+) dramatically outperform plain HTML and JSON-LD conditions. C6+ achieves the highest scores. Error bars show 95% confidence intervals.

Cond.	Answer Summary	Acc.
Factual — “What is Restaurant im Hotel Zauchenseehof? Describe its key features.”
C1	Generic description without specifics. No cuisine type, location, or opening hours mentioned.	1
C2	Identifies it as a FoodEstablishment with some Schema.org properties.	3
C4	Agent searches but finds limited structured data to traverse.	2
C6	Agent follows links to related entities (hotel, region); retrieves cuisine, address, coordinates, and related attractions.	5
C6+	Enhanced+ agent instructions and richer linked-entity navigation surface the same information with lower variance; answer matches C6 quality.	5

Table 3: Qualitative comparison of generated answers across conditions for two representative queries. Accuracy scores are assigned by the LLM judge (1–5 scale).

Figure 4: Answer quality progression across conditions for the same factual query. C1 (plain HTML) produces a generic answer (1/5), while C6 and C6+ (enhanced entity pages + agentic RAG) follow links to related entities and retrieve comprehensive structured data (5/5). C6+ achieves the same peak score with lower variance.

Figure 5: Accuracy improvement waterfall showing the cumulative effect of each optimization layer: JSON-LD markup, agentic retrieval, and enhanced entity pages. The largest gains come from link materialization in enhanced pages, not from adding structured data alone.

Hyp.	Metric	$\Delta$	$t$	$p_{\text{adj}}$	$d$	Sig.
Structured data (standard RAG) (C1 vs C2)
	Accuracy	+0.17	-3.12	2.4e-02	0.18
	Completeness	+0.18	-2.86	5.5e-02	0.16	n.s.
Enhanced pages (standard RAG) (C1 vs C3)
	Accuracy	+1.04	-10.53	2.7e-21	0.60
	Completeness	+1.42	-13.00	5.0e-30	0.74

Table 4: Paired

t

-tests with Bonferroni correction (

\alpha=0.05

n_{\text{tests}}=12

Query Type	C1	C2	C3	C4	C5	C6	C6+
Factual	2.74	3.14	4.57	4.06	4.13	4.56	4.81
Relational	4.48	4.55	4.67	4.56	4.75	4.79	4.85
Comparative	4.77	4.94	4.97	4.83	4.69	4.90	4.95

Table 5: Mean accuracy by condition and query type.

Domain	$n$	C1	C2	C3	C4	C5	C6	C6+	$\Delta_{\text{C6+}}$
BlackBriar	137	4.92	4.89	4.91	4.75	4.74	4.96	4.99	+0.07
Expr. Legal	111	3.36	4.20	4.29	4.33	4.29	4.32	4.86	+1.50
SalzburgerLand	79	2.19	2.33	4.92	4.06	4.25	4.82	4.66	+2.47
WordLift Blog	22	1.91	1.73	4.55	3.14	3.38	4.45	4.64	+2.73
All domains	349	3.62	3.89	4.69	4.36	4.40	4.70	4.85	+1.10

Table 6: Mean accuracy by condition and domain, with per-domain effect sizes for the key C6+ vs. C1 comparison.

n

= number of queries per domain.

Figure 6: Mean accuracy by domain and condition. The direction of the effect is consistent across all four verticals, but its magnitude varies with domain characteristics—largest where the knowledge graph provides information beyond the plain HTML (travel, editorial), near-zero where the baseline is already fact-rich (e-commerce).

Metric	C4	C5	C6	C6+
Links followed	1.0	0.5	0.5	0.4
Links available	41.7	41.9	77.4	102.2

Table 7: Agentic-specific metrics across conditions.

Figure 7: Heatmap of mean accuracy scores across domains and conditions. Darker cells indicate higher accuracy. The pattern shows that enhanced entity pages (C3, C6) consistently outperform other conditions across all domains.

Retrieval-Augmented Generation

Retrieval-augmented generation connects language models to external knowledge so answers can use current, domain-specific evidence.

See all 766 papers

SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges
Aug 12, 2026Yuchao Wu, Junqin Li, XingCheng Liang +4Multi-Hop QA Retrieval-Augmented Generation
Zleap AI
Do We Still Need GraphRAG? Benchmarking RAG and GraphRAG for Agentic Search Systems
Apr 1, 2026Dongzhe Fan, Zheyi Xue, Siyuan Liu +1Graph Retrieval-Augmented Generation RAG Evaluation
New York University Shanghai, Shanghai, China
Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval
Jun 4, 2026Anuj Maharjan, Devinder Kaur, Richard MolyetRetrieval-Augmented Generation Adaptive RAG
Dept. of Electrical Engineering and Computer Science University of Toledo Toledo, OH, USA

Aug 12, 2026J/K move · Enter open · S save

SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges

Yuchao Wu, Junqin Li, XingCheng Liang, Yongjie Chen+3

Zleap AI

Multi-Hop QA Retrieval-Augmented Generation Graph-Based Retrieval

Open paper AskPDF arXiv

While retrieval-augmented generation (RAG) has proven effective at giving LLMs access to external knowledge, mainstream dense-retrieval implementations remain inherently limited in handling structured constraints and multi-hop reasoning. Graph-based methods address this by constructing knowledge graphs offline, but they often fragment semantics, incur high maintenance, and complicate incremental updates. We propose SAG (SQL-Retrieval Augmented Generation), a structured retrieval architecture that organizes documents into an event-entity index without building a global knowledge graph. SAG represents each chunk as a semantically complete event paired with its entities, forming a latent hyperedge that preserves n-ary relations without decomposing them into triples. At query time, SAG treats shared entities as join keys to connect related chunks. This dynamically yields a query-scoped neighborhood of events, and yet every piece of evidence remains the original chunk throughout. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show that SAG achieves the best retrieval and end-to-end QA performance on every benchmark, with gains that widen as reasoning-chain complexity increases. On MuSiQue, where multi-hop evidence chaining is most demanding, SAG reaches 80.36% Recall@5, outperforming the strongest baseline by 11.52 points. This work paves the way for knowledge infrastructure that enables LLM agents to retrieve and reason over continually growing organizational knowledge.

Abstract

Retrieval-Augmented Generation

Explore similar work

SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges

Do We Still Need GraphRAG? Benchmarking RAG and GraphRAG for Agentic Search Systems

Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval