Voice of customer extraction from product reviews: a benchmark of RAG variants

This study evaluates the effectiveness of Generative Artificial Intelligence (G-AI) models enhanced with Retrieval-Augmented Generation (RAG) for automating Voice of Customer (VOC) creation. Four Generative AI architectures were compared using product reviews as the dataset: (1) baseline large language model without retrieval, (2) RAG model with feature labeling, (3) Self-RAG with feature labeling and (4) Sentiment Aware Self-RAG with feature labeling. Models were evaluated across six dimensions: requirements to Critical to Quality (CTQ) coherence, CTQ measurability, description representativeness, topic coverage, desiderata to requirement consistency and overall performance. Sentiment aware Self-RAG model and Self-RAG model with structured feature labeling demonstrated superior performances in generating consistent and comprehensive VOC insights. The Sentiment Aware Self-RAG is an innovative retrieval-augmented strategy that incorporates both semantic similarity and sentiment signals, enabling a more context sensitive generation of VOC insights. Results highlight the potential of Sentiment Aware and feature driven retrieval strategies to improve both the consistency and the depth of VOC generation, providing a more robust foundation for product innovation and customer-centric decision-making. By bridging methods from informatics and marketing, the paper contributes to the development of Artificial Intelligence (AI) driven approaches that enhance the translation of customer voices into actionable product requirements.