SME Business Intelligence Support Using Retrieval-Augmented Generation and RFM Segmentation

Abstract

This study proposes and empirically evaluates an integrated SME business intelligence support system that combines Retrieval-Augmented Generation (RAG) with Recency–Frequency–Monetary (RFM) customer segmentation and embeds both capabilities directly into a mobile keyboard interface for everyday business communication. Unlike conventional chatbots or standalone analytics tools, the system delivers knowledge-grounded automated responses and actionable customer insights within the seller’s existing messaging workflow, eliminating the need for separate applications, local infrastructure, or AI expertise. The framework constructs a structured SME knowledge base in Markdown, applies semantic chunking and Voyage-3 embeddings, and performs vector retrieval via PgVector to ensure high-fidelity grounding before generation using a cloud-based LLM. In parallel, historical invoice data are processed through an RFM engine to classify customers into Loyal, Moderate, and At-Risk segments for targeted promotions. Using real SME data collected over several weeks, the system was evaluated through retrieval faithfulness testing, correctness analysis with confidence intervals, silhouette validation of clusters, end-to-end latency measurement, and User Acceptance Testing with 18 sellers. Results show very high retrieval faithfulness (0.997), strong generative correctness (0.88), acceptable real-time latency (~5 seconds), and stable segmentation performance (Silhouette 0.61; ROC–AUC: At-Risk 0.93, Loyal 0.85). The key novelty lies in unifying RAG-based conversational support and lightweight customer analytics inside a keyboard-level interface, creating a practical, low-barrier pathway for AI adoption in small businesses while preserving natural communication practices.