Implementation and Application of a Generative AI Virtual Librarian
Yun-Fan Chen, Hanjun Ma · International Journal of Librarianship · 2025
Since its transformation into Taiwan’s first national digital library in 2013, the National Library of Public Information (NLPI) has continuously adopted innovative technologies to advance intelligent services. In response to the rise of large language models (LLMs), NLPI launched the “Generative AI Virtual Librarian” project in 2023 and developed Xiaoshu, a virtual librarian capable of natural voice interaction. Centered on generative AI, the system integrates four databases and eight intent categories, applying retrieval-augmented generation (RAG) and speech recognition to provide collection search, book recommendations, service inquiries, and social interaction. Xiaoshu effectively handles a large number of repetitive yet diverse library inquiries, demonstrating the linguistic flexibility of generative AI. Compared with rule-based systems that rely on extensive pre-set Q&A pairs, generative AI reduces maintenance workload by shifting the focus from data quantity to data quality. From late 2023 to October 2025, Xiaoshu recorded over 70,000 interactions, serving about 3,000 users monthly with an accuracy rate above 80%, and reducing human librarian workloads by approximately 16 hours per month. The project highlights the importance of defining clear service goals, user scenarios, and resource planning in the early stages. Overall, NLPI’s experience shows that generative AI librarians can enhance service efficiency and create a new paradigm for human–AI collaboration in public libraries.