Optimizing Retrieval Components for a Shared Backbone via Component-Wise Multi-Stage Training

TL;DR

This paper introduces a multi-stage training framework for optimizing shared dense retrieval backbones in industrial legal systems, improving retrieval quality and system flexibility.

Contribution

It proposes a component-wise, mixed-stage configuration approach for dense retrievers and rerankers, enhancing performance in shared retrieval backbones.

Findings

Different retrieval components show stage-dependent trade-offs.

Component-wise, mixed-stage configurations outperform single checkpoint setups.

The optimized backbone supports multiple applications effectively.

Abstract

Recent advances in embedding-based retrieval have enabled dense retrievers to serve as core infrastructure in many industrial systems, where a single retrieval backbone is often shared across multiple downstream applications. In such settings, retrieval quality directly constrains system performance and extensibility, while coupling model selection, deployment, and rollback decisions across applications. In this paper, we present empirical findings and a system-level solution for optimizing retrieval components deployed as a shared backbone in production legal retrieval systems. We adopt a multi-stage optimization framework for dense retrievers and rerankers, and show that different retrieval components exhibit stage-dependent trade-offs. These observations motivate a component-wise, mixed-stage configuration rather than relying on a single uniformly optimal checkpoint. The resulting backbone is validated through end-to-end evaluation and deployed as a shared retrieval service supporting multiple industrial applications.