Legal text retrieval in multilingual contexts, such as those found within Ecuadorian judicial environments, presents significant challenges due to specialized terminology and inherent semantic complexity. This research proposes a methodological framework, currently at an initial doctoral research stage, designed to refine and specialize pre-trained multilingual embeddings specifically for legal text retrieval tasks. By integrating a Mixture-of-Experts (MoE) architecture with contrastive learning techniques applied explicitly at the embedding level, the proposed approach aims to enhance semantic alignment, embedding uniformity, and domain specialization. This integration specifically addresses recognized limitations of multilingual embeddings such as semantic anisotropy and insufficient domain adaptation. Future empirical validations will be conducted using standard retrieval metrics—including Normalized Discounted Cumulative Gain (nDCG@10), Mean Reciprocal Rank (MRR), and Spearman correlation—to rigorously assess anticipated improvements over existing baseline methods.