<p>Deep learning has revolutionised the field of image classification, enabling outstanding performance in variety of applications. The transformer model was not used for image classification but the emergence of vision transformer (VIT) has resulted in excellent performance in visual domain. Many VIT variants have been developed to improve various components of attention mechanisms. This paper introduces Geometry-Aware attention mechanism for Vision Transformers, replacing the traditional dot-product attention with a geometry-aware formulation. Instead of depending only on inner products, compute the angular distance between query and key vectors and use it to adaptively shift the key vectors either toward or away from the queries. This shift is guided by a piecewise function, which allows the model to enhance or suppress attention based on directional similarity. The method is theoretically grounded in vector geometry, providing a more interpretable and discriminative attention mechanism. We validate our model on CIFAR-10, CIFAR-100, MNIST, and the Citrus Leaves dataset. Compared to the standard ViT, our model achieves a 3% increase in accuracy on CIFAR-10, 2 % on CIFAR-100, 1% on MNIST, 0.6 % on Fashion MNIST and a significant 3 % improvement on the Citrus disease dataset with reduced computational cost and high convergence rate.</p>