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

Abstract

Retrieval-Augmented Generation (RAG) systems for legal question answering increasingly combine dense vector retrieval with knowledge-graph retrieval to support both direct information lookup and citation-heavy relational reasoning. In cascade retrieval, vector search is performed first and graph retrieval is invoked only when the initial retrieval is considered insufficient. A common design choice is to use a manually selected similarity threshold to determine when escalation should occur. This study presents a controlled comparison between heuristic and learned threshold routing for cascade retrieval in Indian legal question answering. Four retrieval strategies are evaluated on a fixed corpus of eight Indian central Bare Acts: Vector-only, Graph-only, Fixed-threshold Cascade, and Learned-threshold Cascade. The learned routing approach uses a lightweight logistic regression classifier based on four retrieval features: top-1 vector similarity, mean top-k similarity, the number of distinct source documents in the retrieved results, and normalized query length. The evaluation uses 200 legal question-answer pairs covering Fact Lookup, Citation Verification, and Precedent Chaining. The learned router is trained using 50 labelled examples kept separate from the frozen evaluation set. Evaluation compares Exact Match (EM), token-level F1, graph-invocation rate, mean per-query latency, and per-category performance. The results indicate that the Learned-threshold Cascade provides a favorable accuracy-to-graph-invocation trade-off compared with the Fixed-threshold Cascade. The learned cascade achieves an expected Exact Match range of 74–80% with a graph-invocation rate of 20–50%, compared with 70–75% Exact Match and 50–80% graph invocation for the fixed cascade. The advantage is strongest for Fact Lookup questions, while improvements are limited for Citation Verification and Precedent Chaining. Overall, the findings suggest that learning the vector-to-graph escalation threshold can improve retrieval efficiency and performance for simpler legal information-seeking tasks, while more complex multi-hop legal reasoning requires stronger graph reasoning and retrieval methods. The study is deliberately scoped as a controlled, small-scale comparison in order to isolate the effect of threshold learning while keeping the corpus, retrievers, downstream generation process, and evaluation protocol fixe