Learned vs. Heuristic Thresholds for Cascade Retrieval Routing in Indian Legal Question Answering

Retrieval-Augmented Generation (RAG) systems for legal question answering commonly combine dense vector retrieval with knowledge-graph retrieval through cascade routing, where graph retrieval is invoked only when vector retrieval confidence falls below a predefined threshold. Existing systems generally rely on manually selected heuristic thresholds, with limited empirical justification. This work presents a controlled comparison between heuristic and learned threshold routing strategies for Indian legal question answering. Four retrieval strategies—Vector-only, Graph-only, Fixed-threshold Cascade, and Learned-threshold Cascade—are evaluated on an Indian legal corpus consisting of Bare Acts and legal judgments. The learned routing approach employs a lightweight logistic regression classifier trained on retrieval features to determine when graph retrieval should be invoked. Experimental evaluation compares retrieval accuracy, Exact Match (EM), token-level F1, graph invocation rate, and latency. Results indicate that learned threshold routing improves retrieval performance for straightforward legal fact lookup tasks while offering limited gains for more complex citation verification and precedent chaining queries. The study provides an empirical analysis of adaptive retrieval routing for legal RAG systems and highlights directions for future research in graph-based legal reasoning.