LLM-Powered Root Cause Intelligence for Inventory and Shipment Discrepancies in Warehouse Management Systems

Inventory and shipment discrepancies in Warehouse Management Systems (WMS) impose significant operational costs on global supply chains, yet current diagnostic approaches remain fragmented, reactive, and heavily reliant on domain-expert intervention. This paper proposes the LLM-Powered Root Cause Intelligence (LRCI) framework, a novel architecture that unifies a fine-tuned Large Language Model (LLM) with Retrieval-Augmented Generation (RAG), a Graph-based Causality Engine (GCE), and a structured anomaly detection pipeline for real-time, interpretable root cause analysis of WMS discrepancies. The framework ingests multi-source operational data — WMS event logs, ERP transaction records, IoT sensor telemetry, and carrier API feeds — and processes them through a four-stage pipeline culminating in LLM-generated root cause diagnoses with natural-language explanations and remediation recommendations. Evaluated on an industrial dataset of 18,500 discrepancy records across six distribution centres, the proposed LRCI framework achieves 94.7% accuracy, 93.1% precision, 95.2% recall, and an F1-score of 94.1%, with a mean root cause identification latency of 1.8 seconds. These results represent consistent improvements of 4.6 to 23.4 percentage points over rule-based, machine-learning, and zero-shot LLM baselines, demonstrating the framework's practical viability for real-time deployment in high-volume warehouse environments.