Akademik Veri Yönetim Sistemi
11th Intelligent Systems Conference, IntelliSys 2025, Amsterdam, Hollanda, 28 - 29 Ağustos 2025, cilt.1567 LNNS, ss.142-160, (Tam Metin Bildiri)
Retail operations create vast amounts of raw data that could give many insights about the work for further restocking optimization. This paper presents a framework that combines Retrieval-Augmented Generation (RAG) and Chain of Thought (CoT) reasoning with Large Language Models (LLMs) to enhance restocking optimization. The proposed approach orchestrates multiple analytical sub-tasks; including data cleaning, feature engineering, outlier detection, correlation analysis, dimensionality reduction, clustering, embedding-based product description analysis, and time-series analysis. Each step generates structured insights, which are sequentially connected using the CoT methodology. The RAG framework leverages these insights as contextual inputs, allowing the LLM to synthesize information and produce interpretable, data-driven recommendations. This process mimics a real-world approach, where specialists from divergent disciplines feed the supervisor from their perspective, and before concluding the process with a final decision, all related information is considered by the supervisor. Similarly, here the supervisor is the LLM, specialists are insight generator RAG models, and the blackboard is the CoT approach. The final decision-making step creates demand forecasting for the prioritized clusters to optimize restocking strategies guided by the LLM, aiming to offer detailed recommendations for high-priority products. This framework provides a scalable solution for retail analytics with the help of real world data and addresses common limitations in conventional inventory forecasting, such as isolated analysis and the lack of interpretability. By bridging structured data analysis with generative AI reasoning, this study demonstrates the significant potential of combining RAG and CoT methodologies for adaptive and explainable retail decision-making frameworks.