The growth of big data across domains poses challenges to information retrieval (IR) systems, especially for traditional supervised methods that rely on costly labeled datasets and limit scalability. This paper introduces a novel self-supervised learning (SSL) algorithm designed for unsupervised IR in large-scale data environments. The algorithm combines contrastive learning and generative modeling to learn semantically rich representations without labeled data. Contrasting learning learns differences between similar and dissimilar examples by using augmented pairs, while generative modeling captures intrinsic properties of data using masked language models and image reconstruction tasks. A unified optimization strategy that combines these methods enables robust performance on a wide variety of data.