Legal documents by nature are complicated, long, sequential in structure, and specialized in terminology. Their subtle context and legal semantics required for a proper interpretation are often lost in traditional summarization methods. The work explores the effectiveness of different word embedding and transformer models—Word2Vec, GloVe, BERT, T5 (Text-to-Text Transfer Transformer), and PEGASUS—for legal document summarization. Our work emphasizes the compromises between abstractive and extractive summarization models in the legal domain. We generate domain-specific embeddings by fine-tuning and training these models on a dataset of Indian Supreme Court judgments. The aim is to produce abstractive summaries that are able to preserve key legal facts while enhancing readability and removing redundancy. The summarization is assessed through semantic similarity metrics and recall-based measures such as ROUGE-1, ROUGE-2, and ROUGE-L [18]. Experimental outcomes reveal that GloVe embeddings, when fine-tuned from legal documents, perform better in extractive summarization with a ROUGE score of 0.07–0.08. Although transformer-based models such as T5 and PEGASUS produce more coherent and fluent summaries, they have lower lexical overlap with reference texts. Our results show that conventional embeddings, when optimized for the legal context, are capable of achieving competitive and effective solutions for context-sensitive legal text summarization and are hence applicable to practical legal information retrieval systems.