May 2026 Summaries

Zilliz's introduction of Vector Lakebase marks its evolution from a vector database system to a unified, lake-native data platform designed for AI workloads. This transition does not signify a departure from vector databases but instead represents the next stage in their development, addressing the limitations of existing architectures by integrating retrieval and large-scale discovery into one operational system. Vector Lakebase combines the semantic retrieval strengths of vector databases with the storage efficiency and analytical capabilities of data lakes, allowing enterprises to handle unstructured data more iteratively and efficiently. By incorporating storage-compute separation, multi-layer caching, and various compute modes, Vector Lakebase aims to provide a cohesive infrastructure that supports both online serving and offline discovery processes. This new architecture addresses the growing complexity and demands of AI systems, ensuring that improvements in data quality and retrieval continuously feed back into production, thus transforming unstructured data management into a continuous operational loop.

Zilliz Cloud On-Demand Search is a new feature introduced to address specific workload needs that are not effectively met by existing Dedicated or Serverless options, particularly in scenarios involving sparse and bursty analytics. The feature emerged from a case study with an autonomous-driving customer whose analytics workload, requiring vector search on a 1 billion-row collection, was inefficiently serviced by both Dedicated and Serverless clusters due to their pricing and operational models. On-Demand Search offers a solution by billing per minute of actual compute usage, loading only the necessary data for each query, and providing workload isolation through separate compute resource groups. This approach reduces costs significantly for workloads with sporadic and unpredictable access patterns, bringing the monthly bill under $500 compared to the much higher costs of the other models. However, On-Demand is not suitable for high-QPS scenarios due to its reliance on IVF indexes and potential latency in cold-query situations. It is part of the Zilliz Vector Lakebase architecture, which supports various AI workloads by enabling different compute shapes to access the same data without copying or syncing, thus optimizing resource use and cost.

Vector Lakebase emerges as a novel architectural solution to address the challenges posed by data gravity in AI systems, where traditional architectures lead to data duplication and synchronization burdens. This new paradigm integrates the capabilities of vector databases with data lakes, offering a unified layer that eliminates the need for separate systems and data movement. By storing and managing AI data, vectors, and indexes directly in object storage, Vector Lakebase enables both online and offline AI operations to share the same source of truth, thereby reducing the operational overhead associated with data migration and synchronization. The system's design supports high-performance, low-latency vector searches and cost-efficient batch processing, making it suitable for a wide range of AI workloads including real-time recommendations, agent memory management, and context engineering. This approach not only accelerates AI feature development but also aligns with the industry's shift towards integrating AI-native operations within existing data infrastructures, as exemplified by Zilliz's public preview of Vector Lakebase.

Zilliz spent eight years optimizing vector databases for faster and more predictable search capabilities, but the focus has now shifted to balancing performance with cost-effectiveness in response to evolving user needs driven by AI advancements. The introduction of Zilliz Vector Lakebase addresses the inefficiencies of always-on compute for data that is infrequently accessed by allowing semantic data to persist independently of a continuous serving cluster. This approach provides on-demand computation that supports multiple compute lifecycles, offering a more dynamic cost model for workloads that are mostly idle. The system leverages innovations such as quantization for reduced cold start times, IVF clustering to minimize data scanning, and advanced storage formats to alleviate I/O amplification, resulting in a scalable, flexible, and economic solution that aligns with the demands of modern AI applications.

Zilliz has introduced Vector Lakebase, an advanced semantic-centric data platform that extends beyond traditional vector databases to support AI workloads by integrating open storage and elastic compute. Built on an S3-based unified data foundation, Vector Lakebase caters to real-time retrieval, iterative discovery, and batch analytics, allowing seamless scaling from gigabytes to petabytes. Unlike vector databases that primarily facilitate real-time serving, Vector Lakebase provides a comprehensive solution by unifying various data types—raw multimodal, semantic, and feedback data—into a structured data plane that supports the continuous loop of AI systems, including serving, learning, and improving processes. It offers features such as tiered serving solutions, on-demand search, external data lake search, full-spectrum search, and unified lake-native storage, addressing challenges like fragmented data architecture and isolated infrastructure that can hinder AI development. With its focus on efficient I/O and advanced search capabilities across dense and sparse vectors, text, JSON, and geospatial data, Vector Lakebase accelerates AI development in diverse application scenarios, from real-time serving and iterative discovery to batch analytics, supporting both existing and evolving data models.