Tacit Knowledge Embeddings and Quantitative-Qualitative Information Fusion for Stockout Prediction and Emergency Ordering Decision Support in Automotive Supply Chains

Demand, inventory and scheduled-receipt figures may not fully reveal supply-chain conditions requiring intervention. This study proposes a decision-support method that uses tacit knowledge embeddings and quantitative-qualitative information fusion to predict stockouts and guide emergency ordering in automotive supply chains. A corpus of 2,128 experience records curated from classroom discussions is consolidated into 365 documents. An embedding model retrieves relevant experience, predefined mappings connect qualitative evidence to numerical assumptions, and a calculation module predicts shortages and selects order quantity, timing and supply route. A private language model composes explanations from retrieved evidence and calculated facts. Six situations, two query formulations per situation and five random seeds yield 3,000 base trajectories, evaluated under four experience non-realization rates and six policies for 72,000 simulation rollouts. At a 25% non-realization rate, experience-based retrieval-augmented generation reduces mean shortage from 128.93 to 19.73 units and total cost by 72.3% relative to a numerical baseline. It does not outperform a stronger baseline supplied with the marginal risk distribution. When none of the anticipated experience-based disruptions occurs, its cost advantage over the numerical baseline reverses. The results support complementing quantitative assessment with relevant qualitative experience, while showing that procurement outcomes depend on retrieval accuracy and the applicability of the retrieved knowledge. The accompanying ZIP contains only experiment reproduction materials: input data, experiment code, fixed model identifiers, dependencies, recorded results, verification scripts, figure-regeneration code, and reproduction instructions.