A Framework for Finished Goods Safety Stock Estimation Using Local LLM-Based Retrieval-Augmented Generation (RAG)

This study proposes a local LLM-RAG-based framework for estimating finished-goods safety stock in high-mix low-volume manufacturing environments. The proposed framework maintains a statistical safety stock calculation as a quantitative anchor and complements it with context-aware reasoning using a local Large Language Model (LLM) and Retrieval-Augmented Generation (RAG). The framework first preprocesses ERP/MES data through SQL-based filtering and calculates dynamic indicators such as risk lead time, seasonality, urgent-order volatility, demand trends, and open orders. These indicators are then injected into the local LLM as structured RAG context to generate explainable safety stock adjustment rationales. The framework was implemented using empirical data from an automotive parts supplier and applied to 180 finished-goods items. The developed program generated safety stock output logs over 22 ASOF points from January to May 2026. Because safety stock is a policy-level reference value linked to production planning and MRP, the quantitative evaluation was conducted using five monthly closing points. The results showed positive penalty improvement in all monthly closing periods, while the delivery response proxy increased from 80.7% to 83.0%. These findings indicate that the proposed framework can support context-aware safety stock adjustment by jointly considering ABC grade, seasonality, demand trends, urgent-order volatility, and open orders, while avoiding excessive weekly fluctuations in policy-level safety stock values.