Bilingual Hybrid Judicial-Opinion Retrieval for Puerto Rico and Texas: Dense, Lexical, and LLM-Augmented

Abstract

This article presents the design, implementation, and evaluation of a bilingual English/Spanish information retrieval prototype over judicial opinions from Puerto Rico and Texas courts. Judicial opinions are ingested from the CourtListener REST API, cleaned, segmented into paragraph-aware chunks, embedded with a multilingual sentence-transformer model, and indexed in Facebook AI Similarity Search (FAISS) for dense retrieval, with a parallel Best Matching 25 (BM25) index supporting lexical retrieval. A weighted Reciprocal Rank Fusion (RRF) layer combines semantic and lexical rankings. An optional layer adds query rewriting and cross-encoder reranking using a locally hosted Gemma 3 large language model. On a stratified held-out split of thirty-two queries, weighted hybrid RRF reaches a normalized discounted cumulative gain at rank 10 (nDCG@10) of 0.5615, while cross-encoder reranking raises this value to 0.7680 at roughly twenty times higher latency. The system preserves end-to-end traceability to source judicial opinions and is presented as a research prototype, not a legal-advice product.