Abstract

Dense retrieval has become a critical component of modern knowledge-intensive AI systems, enabling efficient semantic search over large document collections for applications such as question answering, enterprise search, recommendation, and retrieval-augmented generation. In practice, however, many high-value corpora are inherently distributed across organizations, regions, devices, or business units and cannot be centralized due to privacy, compliance, ownership, or cost constraints. Federated Learning (FL) provides a natural paradigm for collaborative model training without sharing raw data, yet applying FL to retrieval systems introduces a fundamental challenge: embedding inconsistency across clients under heterogeneous knowledge distributions. When local retrievers are trained on domain-skewed corpora and aggregated natively, the resulting representation space can drift across clients, leading to degraded ranking quality, unstable convergence, and poor cross-client interoperability. This paper presents Federated Retrieval Alignment (FRA), a framework for training dense retrieval models in federated environments while preserving representation consistency across heterogeneous clients. FRA augments the standard retrieval objective with an alignment-aware regularization term that constrains local query embeddings toward a shared global embedding geometry, thereby reducing drift introduced by heterogeneous document and query distributions. We implement a prototype using dual-encoder retrieval models and evaluate it on partitioned Natural Questions and TriviaQA workloads with controlled domain skew, query skew, and client imbalance. Experimental results show that FRA improves Recall@10 by up to 9-12% relative over standard Federated Averaging under heterogeneous settings, reduces embedding drift by more than 60%, and stabilizes convergence without introducing additional communication overhead. Ablation studies further demonstrate that moderate alignment strength provides the best trade-off between retrieval quality and optimization speed. These findings indicate that embedding-aware coordination is essential for reliable federated retrieval learning and that native federated aggregation is insufficient for distributed knowledge systems operating under realistic heterogeneity.