To address the limitation that existing received signal strength (RSS)-based indoor localization methods rely on extensive labeled information for feature extraction in complex indoor electromagnetic transmission environments, this study proposes a few-shot indoor localization method based on contrastive learning. The method treats radio frequency (RF) fingerprints from the same location as positive samples, while fingerprints from different locations are considered negative samples. The model can learn an efficient distance metric for distinguishing RF fingerprints from different locations, thus enabling more effective extraction of RF fingerprint features. In addition, a penalty term related to physical distance, namely spatial-aware loss, is designed during feature extraction network training to ensure that the model adequately considers physical spatial constraints, so that spatially proximate samples are also closer in the feature space, resulting in a robust feature extractor. With the pre-trained feature extractor kept unchanged, a multi-layer perceptron (MLP) is trained to learn the mapping relationship between extracted features and corresponding location coordinates to form a terminal position localization network. Simulation experimental results show that the proposed method achieves a performance improvement of approximately 5% compared with the localization method combining phase calibration and SSIM data augmentation.