LibraAI: a lightweight conversational AI system for book consultation in e-commerce

Dinh, Bao-Minh, Nguyen, Truong-Son, Trinh Thi Mai, Huong, Phan, Thanh-Duy, HoangVan, Xiem
Electronic Commerce Research · Springer Nature · 2026-10-08 · DOI: 10.1007/s10660-026-10203-7

Аннотация

Conversational AI systems are increasingly common on e-commerce platforms, where they provide instant product consultation and personalized recommendations. Graph-based retrieval-augmented generation (RAG) frameworks have further advanced these systems by enabling more natural and contextually aware responses. In this paper, we present LibraAI: a lightweight conversational AI designed to help users find suitable books on Tiki, a Vietnamese e-commerce platform with a highly structured and specialized book catalog. LibraAI is built on an enhanced graph-based RAG framework and introduces a novel relation-oriented retrieval method. Our experiments show that LibraAI reduces input-context token usage by 67% relative to LightRAG while maintaining comparable response times. Against LightRAG, MiniRAG, and PathRAG, LibraAI matches or exceeds baseline comprehensiveness and empowerment at substantially lower computational cost. Comprehensive evaluations, including both large-language-model analysis and human user surveys, demonstrate that LibraAI consistently outperforms Tiki’s incumbent Assistant in terms of answer comprehensiveness, diversity, and user empowerment. Beyond technical efficiency, LibraAI addresses practical challenges of AI adoption in e-business, such as scalability, adaptability to dynamic markets, and computational cost. This work positions LibraAI as a strong example of sustainable AI integration in digital commerce, with potential for broader application in other product domains and compliance-sensitive services.

Читать у издателя →

Похожие статьи