Year: 2025
Type: dissertation
Abstract: The exponential growth of complex legal documents in India has made it difficult for non-experts to access and interpret statutory information which is mainly in domains such as automobile law. This study addresses that challenge by developing a Retrieval-Augmented Generation (RAG) framework for legal question answering by focusing on the Motor Vehicles Act, 1988. The study aims to identify the most powerful embedding model for accurate semantic retrieval and contextual response generation. There are eight embedding models which were evaluated using metrics such as Recall, Precision, nDCG, F1, ROUGE-L, BLEU, and Cosine Similarity. All experiments in this study were conducted exclusively on English legal text. The dataset was preprocessed, chunked, and vectorized, with embeddings stored in ChromaDB and used via LangChain for retrieval and LLaMA-based LLMs for answer generation. This study results showed that multilingual-e5-large achieved the highest semantic similarity (0.9558) and balanced accuracy across all metrics, outperforming both general-purpose and domain-specific models. The findings theoretically advance the understanding of embedding alignment in legal semantic retrieval and, in practice, offer a scalable framework for automated legal assistance that enhances accessibility and transparency in the justice system.
Author: Yelala, Nithin Reddy
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212349471