Zilliz, the innovator behind Milvus the world’s most widely integrated open-source vector database announced the public preview of Zilliz Vector Lakebase. This major update to Zilliz Cloud integrates a high-performance production vector database with a shared, lake-native data layer.
By keeping Zilliz Cloud’s core real-time vector search capabilities which power enterprises like Zillow, OpenEvidence, Exa, Filevine, and MiniMax and introducing three additional functionalities on the same architecture, the release enables organizations to manage interactive discovery, large-scale batch analytics, and real-time serving on a single data foundation.
The Imperative for a Centralized Data Foundation
Modern AI ecosystems have evolved past isolated, single-query retrieval frameworks. They function as continuous operational loops: serving data, learning from user feedback, extracting insights, preparing optimized data, and serving it again. Traditionally, each phase of this cycle demanded isolated environments for production serving, exploration, and heavy batch processing. Migrating billions of vectors across these disparate infrastructures can take days, creating significant cost and operational friction. As a result, engineering teams frequently bypass the continuous learning loop, leaving crucial corporate data retrievable but stagnant.
Vector Lakebase bridges this infrastructure gap by implementing a zero-copy semantic data plane anchored on shared lake-native storage. This architecture allows batch analytics, interactive discovery, and real-time serving to query one logical copy of the data, scaling seamlessly from gigabytes to petabytes.
“Teams asked for a way to keep their data in one place and run very different workloads against it from real-time agent memory to overnight semantic deduplication,” said Robert Guo, VP of Product at Zilliz and one of the architects behind Milvus. “Vector Lakebase delivers that through a unified storage layer on Vortex, tiered serving for the production path, and on-demand compute for everything else.”
Also Read: Actian Launches Autonomous Data Steward Agent to Guarantee Semantic Consistency Across Enterprise AI Systems
Five Interconnected Capabilities on a Single Architecture
Zilliz Vector Lakebase introduces five core capabilities designed to optimize performance, manage data storage costs, and eliminate data redundancy:
- Tiered Real-Time Serving: To support the various workloads of an enterprise, the platform brings three different performance tiers:
- Performance-optimized: Enables over 1,000 QPS with single-digit millisecond latency through in-memory processing.
- Capacity-optimized: Properly balances performance and scale to give 100 to 500 QPS with latency less than 100 ms, using memory and NVMe.
- Tiered Storage: Reduces costs by using memory, NVMe, and object storage to deliver 10 to 50 QPS with approximately 100ms latency. All tiers ensure 95 to 98% recall (can be configured to exceed 99%) and are backed by a 99. 99% uptime SLA.
- On-Demand Search: Designed for irregular or intermittent workloads where dedicated infrastructure typically idles, this pay-as-you-go model charges strictly for active object storage and processing compute without serverless markups. Internal benchmarks conducted by Zilliz on 1 billion 768-dimension vectors requiring 10 hours of monthly active compute demonstrated a cost reduction to $318, compared to $4,937 on a traditional serverless setup representing approximately 1/15th of the cost.
- External Data Lake Search: Employing a zero-copy “External Collection” mechanism, the platform applies vector indexing and comprehensive search capabilities directly across existing tables in Iceberg, Lance, Parquet, and Vortex formats. Source data remains in its native environment with incremental synchronization upon refresh.
- Full-Spectrum AI Search: The platform supports advanced multi-path retrieval across dense and sparse vectors, text, JSON, and geospatial data. It accommodates hybrid retrieval, BM25 text search, regex, multi-vector queries, and iterative searches. Results can be optimized using modern reranking models like Cohere and Voyage AI, alongside Reciprocal Rank Fusion (RRF) and custom weighted, boost, or decay configurations.
- Unified Lake-Native Storage: Built on Vortex an open columnar format optimized for fast, cost-efficient random reads compared to Lance or Parquet the shared storage layer couples with object-storage-aware vector, BM25, and JSON indexes. This reduces read amplification by over 90%, allowing a 100-million-row schema backfill to finish in single-digit minutes without impacting active operational query traffic.
Availability and Deployment Options
Zilliz Vector Lakebase is available now in public preview on Zilliz Cloud, running alongside the existing Serverless, Dedicated, and Bring Your Own Cloud (BYOC) deployment alternatives across more than 30 cloud regions on Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure. Organizations registering with a valid corporate email address receive $100 in complementary platform credits at the official company website.
“Vector Lakebase is what we believe comes next: one data foundation where the same vectors can serve a production query, anchor a discovery session, and power a multi-petabyte training-data pipeline without copies, migration, or a parallel stack.”
