Hybrid AI-Powered Legal Assistant for BNS Laws Using Structured Retrieval and Contextual Language Model Generation

Abstract: Access to accurate legal information remains a major challenge for non-experts due to the complexity of statutory language and the evolving structure of Indian criminal law. This study proposes a hybrid AI-powered legal assistant that combines deterministic statutory retrieval with contextual large language model generation to provide reliable and interpretable legal explanations. The system integrates a structured legal database containing codified criminal statutes with a contextual explanation engine powered by a locally deployed large language model (LLaMA 3). A dual-layer architecture was implemented: an offline rule-based retrieval module for precise statutory matching and an online generative module for extended legal explanation. Query preprocessing includes keyword expansion, stemming, and multilingual normalization. Performance was evaluated using a curated dataset of 250 legal queries covering direct section lookups and semantic intent-based questions. The offline module achieved high precision for direct statutory queries with low latency suitable for interactive use. The contextual module improved interpretability and semantic coverage for intent-based questions. The hybrid strategy reduced hallucination risk compared to standalone generation by grounding responses in verified statutory content. The proposed hybrid legal assistant provides a scalable and accurate framework for legal information access. By combining deterministic retrieval and contextual generation, the system mitigates hallucination risks associated with large language models while enhancing user comprehension. The framework is designed to be extensible to multilingual and cross-jurisdictional applications.

Index Terms—Artificial Intelligence, Legal Informatics, Hybrid Retrieval, Large Language Models, Indian Criminal Law, Multilingual NLP