Optimizing Retrieval Components for a Shared Backbone via Component-Wise Multi-Stage Training
For practitioners deploying shared retrieval backbones in industrial settings, this work provides a practical solution to optimize component performance across diverse applications, though the findings are domain-specific and incremental.
The paper addresses the challenge of optimizing dense retrieval components shared across multiple industrial applications, proposing a component-wise multi-stage training approach. The resulting backbone improved retrieval quality and was deployed in production legal retrieval systems, supporting multiple downstream tasks.
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.