Abstract:

[Purpose/Significance] University libraries are experiencing a structural change in reference and consultation services: user inquiries are increasingly frequent, fragmented, and cross-disciplinary, while service expectations emphasize immediacy, continuity, and actionable guidance. Under such conditions, large language models may relieve routine workloads and extend service availability, yet their library-specific reliability hinges on whether their responses are grounded in local rules, licensed resources, and auditable evidence. This study examines the deployment of DeepSeek in intelligent QA and service consultation in university libraries, with two goals: 1) to measure its performance across consultation scenarios, disciplinary domains, and inquiry types; and 2) to clarify the mechanisms that explain why effectiveness improves in some settings but not in others. The study differentiates itself from prior discussions by moving beyond general "potential and risk" arguments to a structured evaluation and mechanism-oriented analysis that links local knowledge governance, scenario engineering, human - machine collaboration, and operational constraints to observable service outcomes. [Method/Process] The research adopts a mixed-method design and investigates five university libraries in Henan Province that have piloted or deployed DeepSeek-related services to varying degrees. Data sources include a) user questionnaires capturing usage frequency, task preferences, satisfaction, and perceived value; b) staff questionnaires documenting deployment modes, knowledge-base connections, maintenance routines, quality control practices, and operational challenges; c) in-depth interviews that detail workflow design, escalation rules, and the division of labor between librarians and the system; and d) case materials and system records used for triangulation. In total, 850 questionnaires were distributed and 783 valid responses were collected (625 users and 158 staff). An evaluation framework was constructed along four dimensions - technical performance, service effectiveness, user experience, and managerial benefits - to ensure comparability across libraries. Quantitative analyses include descriptive statistics and group comparisons across inquiry types and disciplines, supplemented by mechanism-oriented interpretation using indicators such as the depth of local knowledge integration, the effectiveness of retrieval augmentation, the degree of scenario customization, and the intensity of governance constraints. Qualitative coding of interviews and case materials was conducted to explain observed differences and to identify operational conditions that enable sustained improvement. [Results/Conclusions] Results show that DeepSeek performs well in routine, rule-based consultations. It substantially improves response timeliness and expands service availability, with an overall satisfaction rate of 86.7%. Deep integration with local library knowledge bases is associated with a marked increase in accuracy for library-specific questions (from 64.3% to 93.7%), improved precision in professional literature recommendations (by 35.2%), and higher efficiency in handling complex academic consultations (by 43.8%). However, effectiveness varies systematically: outcomes are better in science and engineering domains and in factual inquiries than in humanities and social sciences and in research- or innovation-oriented inquiries that require domain judgment and verifiable evidence chains. Mechanism analysis indicates that reliability gains depend on 1) robust local knowledge governance with version control and evidence-first retrieval, 2) scenario-specific templates and graded escalation procedures that standardize outputs by task type, 3) human-machine collaboration that supports librarian review, structured correction, and "write-back" updates, and 4) feedback-driven iteration supported by monitoring metrics and accountable operations. The study also acknowledges limitations: the sample is regionally bounded and may overrepresent early adopters; several measures rely on self-reports and short observation windows; and causal identification is constrained by cross-sectional design and rapid model/version iteration. Future research should expand to multi-region samples, incorporate longer-term operational logs, and employ quasi-experimental designs to strengthen causal inference while addressing privacy, compliance auditing, and sustainable governance in library AI services.

Key words: DeepSeek, library, intelligent QA, service consultation, efficiency improvement, human-machine collaboration