Description As AI applications move from prototypes to production, traditional data infrastructure is starting to show its limits. Systems designed for analytics or transactions struggle to support modern AI workloads — especially vector search, hybrid filtering, and multi-modal data at scale.

In this talk, we present Milvus and the concept of a Vector Lakebase — an AI-native data platform that unifies vector search with lakehouse storage. Instead of building yet another standalone vector database, Vector Lakebase brings indexing and retrieval directly to data lakes such as Iceberg and Paimon.

We’ll dive into the architecture and key design decisions behind Milvus, including:

How to enable efficient vector search directly on Iceberg/Paimon tables without data duplication Zero-copy data access: separating compute and storage while avoiding costly data movement Snapshot-aligned indexing: keeping vector indexes consistent with lakehouse metadata Handling production challenges like multi-tenancy, filtering, and cold-start performance

We’ll also share real-world patterns where this architecture matters most:

Agentic search systems that require consistent, high-recall retrieval Data discovery across large, heterogeneous datasets Data deduplication and similarity-based data management at scale

Finally, we’ll discuss how AI-native data infrastructure is evolving — and why integrating vector search into the lake, rather than building isolated systems, is becoming the dominant pattern.

Speakers:

Xiaofan Luan: Zilliz, Partner and Director of Engineering

Xiaofan Luan, Partner and Director of Engineering of Zilliz, member of Technical Advisory Committee of LF AI & Data Foundation. He worked successively in the US headquarters of Oracle and Hedvig, a software defined storage startup. He joined Alibaba cloud database team and was in charge of the development of NoSQL database HBase and Lindorm. Xiaofan Luan holds a master’s degree in Electronic Computer Engineering from Cornell University.