Pattern identification in documents plays a crucial role in the banking sector and others, as it is applied in various operations such as document analysis, information retrieval, and anomaly detection. Challenges arise when traditional training methods require vast labeled datasets, which are costly and time-consuming. This requirement imposes limitations on the generalization and practical applicability of the models developed. This study evaluates the performance of models in a limited data scenario, by leveraging multimodal information and transfer learning, it aims for accurate identification with reduced data requirements. The ability to train efficient models with scarce data proposes a methodological evolution, reducing the need for large volumes of labeled data and promoting more agile and scalable approaches. We compare Few-Shot Learning (FSL) approaches with reference models. The results demonstrate that, even with limited samples, it is possible to achieve performances comparable to established benchmarks. Future research may further explore the behavior of these methods with varied sample sizes and their ability to detect different types of anomalies in unseen document patterns using the Once Learning approach.