Domain-Partitioned Hybrid RAG for Legal Reasoning: Toward Modular and Explainable Legal AI for India

A team of researchers has introduced a novel AI architecture designed to tackle the complexities of legal research in India. Their paper, "Domain-Partitioned Hybrid RAG for Legal Reasoning: Toward Modular and Explainable Legal AI for India," addresses the failure of standard keyword or embedding-based systems to handle India's long, heterogeneous legal documents spanning statutes, constitutional provisions, penal codes, and judicial precedents. The proposed solution moves beyond standard retrieval-augmented generation (RAG), which struggles with multi-hop reasoning and cross-domain dependencies, by creating a hybrid system that combines domain-specific retrieval with structured relational knowledge.

The core innovation is a system that integrates three specialized RAG pipelines—for Supreme Court case law, statutory texts, and the Indian Penal Code—with a Neo4j-based Legal Knowledge Graph. This graph captures structured relationships among cases, statutes, judges, and citations. An LLM-driven agentic orchestrator dynamically routes queries across these modules, fusing evidence into citation-aware responses. Evaluated on a synthetic 40-question benchmark curated from authoritative sources and assessed via an LLM-as-a-Judge framework, the hybrid architecture achieved a 70% pass rate, substantially outperforming a RAG-only baseline at 37.5%. This demonstrates that combining partitioned retrieval with structured knowledge provides a more scalable and interpretable foundation for advanced legal AI, with significant improvements in reasoning completeness and quality for the Indian judicial context.