A digital design framework for wellness tourism spaces based on local knowledge graphs

Abstract The rapid scaling of wellness tourism in China has produced striking spatial homogenisation, eroding the place-specific cultural texture that distinguishes one region from another. This study proposes a digital design framework that couples local knowledge graphs with parametric generation, and positions it as an addition to the evidence-based, simulation-supported and participatory traditions that already govern spatial design rather than as a replacement for them. Multi-source heterogeneous data—documentary archives, field investigations, geospatial layers, and user-generated content—are assembled into a three-layer graph encompassing concept, instance, and association layers, with provenance preserved across modalities. A semantic mapping mechanism translates retrieved subgraphs into spatial parameters through relational graph attention and learnable projection, while a multi-objective optimisation loop balances cultural fit, spatial suitability, wellness functionality, and user satisfaction. We validated the framework in the Bama Longevity Region, where a gazetteer-enhanced extraction model reached an F1 of 0.871. Two design tracks worked from the same brief, budget, terrain data, software suite and calendar window; under those documented conditions the model-internal cultural-fit score rose from 0.612 to 0.847, a relative gain of 38.4 per cent, spatial use efficiency rose from 64.2 per cent to 78.9 per cent, and the time to a Pareto-acceptable candidate fell from 47.0 to 9.6 h. Because that score is computed in the embedding space that also drives generation, we commissioned an independent blinded panel of 36 heritage specialists, Yao and Zhuang practitioners and resident representatives, which rated the generated scheme 4.80 against 3.65 on a seven-point scale with an ICC(2,k) of 0.85. Structural equation modelling on 286 respondents confirmed that cultural identification mediates restorative experience and overall satisfaction, and comprehensive benefit assessment showed a 0.218 advantage on the normalised composite. The work contributes a fusion strategy, a context-conditioned reasoning mechanism, and a reusable parametric mapping rule set. Since the evidence rests on a single data-rich case and on two teams that differed in composition as well as in method, we present the results as a demonstration of feasibility and report the conditions under which transfer to other regions is likely to weaken.