Traceable Knowledge Graph Reasoning Enables LLM-Assisted Decision Support for Industrial VOCs in the Steel Industry

Researchers developed Chat-ISV, an LLM-enhanced knowledge graph system that organizes fragmented steel industry VOCs literature into a queryable database with 27,180 nodes and 81,779 semantic edges. The system achieved 96.93% precision in answering specialized industrial questions, demonstrating a scalable approach to deploying reliable LLMs in domain-specific applications where hallucination risks are high.

Chat-ISV addresses a critical limitation of general-purpose LLMs: their tendency to hallucinate when answering low-frequency, highly specialized questions. Steel industry volatile organic compounds governance requires precise integration of process knowledge, pollutant data, and control technologies—information scattered across unstructured scientific literature that generic models struggle to synthesize reliably. By constructing a Neo4j knowledge graph from curated literature, the system creates verifiable, traceable connections between concepts rather than relying on probabilistic token prediction.

The architecture combines multiple techniques to enhance reliability: prompt-constrained extraction ensures consistent data parsing, topology optimization reduced isolated knowledge nodes from 57% to 4%, and multi-agent routing directs queries to appropriate knowledge sources. Expert blind evaluations across 400 test cases yielded an F1-score of 0.830 and mean score of 1.69/2.00, validating the approach's effectiveness. Critically, the system enables source-backtracking retrieval, allowing users to trace answers back to original literature—essential for regulatory compliance and liability in industrial contexts.

This work establishes a replicable paradigm for deploying LLMs in specialized domains where accuracy directly impacts operational safety and regulatory compliance. The scalability implications extend beyond steel manufacturing to pharmaceuticals, chemical processing, and aerospace—industries where decision-support tools must maintain audit trails and factual accountability. As organizations increasingly adopt LLMs for knowledge work, knowledge graph augmentation emerges as a production-ready solution for reducing hallucination while maintaining transparency.

• →Knowledge graph-augmented LLMs achieve 96.93% precision on specialized industrial questions, significantly outperforming unstructured model responses.
• →Topology optimization techniques can reduce knowledge graph fragmentation from 57% to 4%, improving information connectivity and retrieval quality.
• →Source-backtracking and traceable reasoning enable regulatory compliance and liability management in safety-critical industrial applications.
• →The Chat-ISV architecture demonstrates a scalable environmental-informatics paradigm applicable across regulated industries beyond steel manufacturing.
• →Multi-agent routing and interactive subgraph visualization enable domain experts to validate and understand LLM-generated decision support recommendations.