Embedding-to-Self-Mined Pipeline Optimizes Industrial Legal Retrieval

For large-scale legal retrieval, an embedding-to-self-mined reranker pipeline offers the best performance-to-complexity trade-off. This strategy outperforms native 4B rerankers by effectively managing hard negatives when scaling from 0.6B to 4B parameter models.