Abstract: The digital transformation of Dunhuang Studies has significantly improved access to manuscripts, paintings, catalogues, and related research materials. Yet the visibility of resources does not automatically translate into their usability for research. Although large-scale digitization projects have substantially alleviated the difficulties caused by the transnational dispersal of Dunhuang materials, core scholarly challenges—such as the aggregation of topic-specific materials, the tracing of documentary provenance, the identification of relationships between versions, the reconstruction of historical contexts, and the establishment of cross-textual knowledge linkages—have not been resolved simply through the enrichment of digital resources. Against this background, this article argues that the next stage of digital Dunhuang research should not be marked merely by the expansion of image repositories or the convenience of catalogue retrieval, but by the construction of knowledge infrastructure oriented toward research practice and characterized by computability and verifiability.

With this objective in mind, the article proposes a localized collaborative research framework for Dunhuang Studies that integrates intelligent document parsing, structured text extraction, knowledge base construction, retrieval-augmented generation, and human verification into a unified workflow. Grounded in the disciplinary characteristics of Dunhuang Studies—namely the dispersal of materials, the coexistence of multiple languages, unstable image quality, complex page layouts, and a strong reliance on provenance awareness and philological control—this framework employs MinerU as the document parsing tool, Qwen3.5 as the locally deployed large language model, and MaxKB as the knowledge base platform. In doing so, it reconstructs a research pipeline that moves from image-based documents to structured knowledge units and then to evidence-driven scholarly interaction. The article emphasizes that the role of artificial intelligence in Dunhuang Studies should not be understood as a substitute for scholarly judgment, but rather as a form of research infrastructure: its function lies in reducing repetitive labor, reorganizing dispersed evidence, supporting exploratory and verification-oriented inquiry, and strengthening the evidentiary basis of historical interpretation.

This article advances three principal arguments. First, the significance of local deployment in Dunhuang Studies lies not only in technical feasibility and data security, but also in its capacity to safeguard scholarly interpretive authority, the continuity of annotation work, and the transparency of evidentiary chains. Second, the core value of retrieval-augmented interaction in this field does not lie in the direct generation of answers, but in the structured reorganization of evidence across manuscripts, catalogues, and prior scholarship. Third, the incorporation of intelligent tools into Dunhuang research should not be interpreted as the automation of humanities scholarship, but rather as a reorganization of the conditions of knowledge production. By facilitating a shift in Dunhuang Studies from “resource visibility” to “knowledge usability,” this collaborative architecture not only offers a new methodological pathway for manuscript studies, but also reveals both the potential and the epistemological limits of AI-assisted humanities research.