Title: Knowledge Graphs: The Essential Data Foundation for LLM-Driven Industrial Asset Management

Abstract:

Current Large Language Model (LLM) agents struggle with accuracy when performing reasoning tasks over unstructured document repositories in industrial asset operations. Data from AssetOpsBench (KDD 2026) reveals that GPT-4 agents score only 65% across 139 industrial maintenance scenarios. While previous studies have compared LLM orchestration frameworks, such as "Agent-As-Tool" versus "Plan-Execute," they relied on a static data layer. This study pivots to investigate the impact of the underlying data model itself.

By employing a typed knowledge graph as a grounding substrate, we route queries through three distinct mechanisms: (i) generating Cypher code via LLMs for structured retrieval, which boosts the GPT-4 performance from 65% to between 82% and 83%; (ii) utilizing native graph optimization primitives without LLM intervention, achieving 99% accuracy on scenarios answerable via graph structures; and (iii) applying Generation-Augmented Knowledge (GAK) for data gaps. In the GAK approach, the agent materializes missing facts as provenance-tagged nodes before providing an answer.

A central finding is the inversion of traditional LLM usage: we restrict the LLM to query generation or one-shot schema enrichment, allowing the graph to handle execution deterministically. When tested against 88 real-world failure-mode scenarios flagged by the benchmark as non-deterministic—specifically involving ten equipment types missing from the initial graph—GAK increased answerability from zero to 100% across all equipment types. It successfully answered 81.8% of these scenarios, with every generated fact clearly marked as "source:LLM-derived" to ensure auditability. Additionally, the study introduces 40 new graph-native scenarios. The results indicate that for structured operational domains, the data layer—not LLM orchestration—is the critical variable, with typed knowledge graphs acting as the essential bridge between raw industrial data and LLM reasoning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC