FedBridge: A Federated Query Engine over Embedding-Heterogeneous Vector Databases

Abstract

Vector databases have emerged as core infrastructure for managing unstructured data by enabling efficient semantic search over high-dimensional embeddings. While existing systems largely assume a centralized setting, real-world deployments are often federated, with data distributed across autonomous silos. A fundamental challenge in such environments is embedding heterogeneity: different silos and query users may adopt distinct embedding models, making their vector representations incomparable and hindering effective retrieval. In this paper, we present FedBridge, a federated query engine designed for embedding-heterogeneous vector databases. FedBridge introduces a lightweight alignment technique to bridge heterogeneous embedding spaces, along with two novel sampling strategies that effectively probe data silos and improve query efficiency. Besides, it provides a unified query interface for embedding-heterogeneous vectors, supports diverse embedding models and data formats, and offers adapters for popular vector database systems. We demonstrate the deployment and usage of FedBridge via a representative scientific question answering scenario.