Companies that are running into performance walls as they scale up their vector databases may want to check out the latest update to Zilliz Cloud, a hosted version of the Milvus database from Zilliz. The database maker says the update brings a 10x boost in throughput and latency, three new search algorithms that improve search accuracy from 70% to 95%, and a new AutoIndexer that eliminates the need to manually configure the database for peak performance on each data set.
Interest in vector databases is booming at the moment, thanks in large part to the explosion in use of large language models (LLMs) to create human-like interactions, as well as increasing adoption of AI search. By caching relevant documents as vectorized embeddings in a database, a vector database can feed more relevant data into AI models (or return better results in a search), thereby lowering the frequency of hallucinations and creating a better overall customer experience.
Zilliz is among the vector databases riding the GenAI wave. As the commercial outfit behind the open source Milvus database, the Redwood City, California company is actively working to carve out the high-end segment of the vector database market. Zilliz CEO and Founder Charles Xie says the company has more than 10,000 enterprise users, and counts large enterprises like Walmart, Target, Salesforce, Intuit, Fidelity, Nvidia, IBM, PayPal, and Roblox as customers.
With today’s update to Zilliz Cloud, customers will be able to push the size and performance of their vector databases installations even more. According to Xie, customers can use the 10x performance boost to either increase the throughput or to lower the latency.
“A lot of these vector database are running queries at subsecond latency,” Xie tells BigDATAwire. “They’re running somewhere from one second to 500 milliseconds. But in terms of latency, a lot of customers may expect more real-time latency. They want the query to be running in milliseconds, basically in tens of milliseconds. They want to get the results in 10 milliseconds or in 20 milliseconds.”
Customers that need more throughput can configure the database to boost throughput. According to Xie, vector databases often deliver to 50 to 100 queries per second. With the update to Zilliz Cloud, the company is able to offer a lot more, Xie says.
“There are a lot of these online services, they want 10,000 queries per second,” he says. “If you get a super popular application, you get hundreds of millions of users, you’d probably like somewhere from 10,000 per second to even 30,000 per second. With our new release, we can support up to 50,000 queries per second.”
The performance boost comes from work Zilliz has done to expand support for parallel processor deployments. It also added support for ARM CPU deployments, to go along with its previous support for Intel and AMD CPUs and Nvidia GPUs. It’s currently working with AWS to support its ARM-based Graviton processors, Xie says.
“We are using the parallel processing instruction set of modern processors, either the ARM CPU or Intel CPU, to unlock the full potential of the parallel data execution,” Xie says.
As companies move GenAI applications from development to production, the size of their vector databases is increasing. A year ago, many vector databases had on the order of a million vector embeddings, Xie says. But at the beginning of 2023, it was becoming more common to see databases storing 100 million to several billion vectors, Xie says. Zilliz’ largest deployment currently supports 100 billion vectors, he says.
Zilliz Cloud customers will be able to get more use out of all that high-dimensional data with the addition of new search algorithms. In previous release, Zilliz Cloud supported dense vector search, including approximate nearest neighbor (ANN). Now it sports four.
“We introduced a sparse index search, or basic sparse embedding search. And we also introduced scalar search, so you can do data filtering on top of a scalar property. And also we have this multi-vector search, so basically you can put a number of vectors in a vector array, to get more context in this search,” Xie explains.
“So combining these four searches–dense vector search, sparse vector search, scalar search, and also multi-vector search–we can bring the accuracy of the search result to another level, from around 70% to 80% accuracy to 95% and above in terms of recall accuracy,” he continues. “That’s huge.”
All those new search types could add a lot more complexity to Zilliz Cloud, further putting the database out of reach of organizations that can’t afford an army of adminstrators. But thanks to the new AutoIndexer added with this release, customers don’t have to worry about getting 500 to 1,000 parameters just right to get optimal performance, because the product will automatically set configurations for the user.
“A vector database is a very complex because it’s basically managing high-dimensional data. There are a lot of parameters and configurations and so the challenges are that a lot of our customers have to hire a bunch of vector database administrators to do all this configuration, to have a lot of trial and error and difficult configurations to get the best configuration for their usage pattern for their workload,” Xie says.
“But with AutoIndex, they don’t need that anymore,” he continues. “It’s autonomous driving mode. We’re using AI algorithms behind the scene to make sure that you get the best configuration out of the box. And the other thing that it also it also beneficial for them to reduce the total cost of ownership.”
A year ago, it was common for customers to spend $10,000 to $20,000 per month on a vector database solution. But as data volumes increase, they find themselves spending upwards of $1 million a month. “They’re definitely looking for a solution that can provide a better total cost of ownership,” he says. “So that’s why cost reduction is been very important to them.”
Zilliz Cloud is available on AWS, Microsoft Azure, and Google Cloud. For more information, see www.zilliz.com.
Related Items:
Zilliz Unveils Game-Changing Features for Vector Search
How Real-Time Vector Search Can Be a Game-Changer Across Industries
November 27, 2024
- NTT DATA Announces Intent to Acquire Niveus Solutions to Accelerate Global Leadership in Google Cloud
- Abu Dhabi’s Technology Innovation Institute Inaugurates Open-Source AI Summit with Critical Discussions on the Future of AI
- IDC Report: AI Infrastructure Spending to Surpass $100B by 2028
- Perplexity Leverages Quartr API to Power First-of-its-Kind Analysis Tools for Retail Investors
November 26, 2024
- Alation Named a Leader in IDC MarketScape: Worldwide Data Intelligence Platform Software 2024
- data.org Launches Asia Pacific Data Capacity Accelerator
- Cisco Reports Decline in AI Readiness Among Canadian Organizations
- LTIMindtree and Microsoft Partner to Supercharge AI Innovation and Digital Transformation for Global Enterprises
- Arcee AI Signs Strategic Collaboration Agreement with AWS to Accelerate the Deployment of Smaller, Specialized Language Models
November 25, 2024
- AI and Analytics Firm datakulture Embarks on Fresh Data Strategy to Drive Growth
- H2O.ai Unveils Agentic AI That Converges Generative and Predictive AI with Purpose-built SLMs
- Elastic Now Collaborates with AWS to Leverage Generative AI Capabilities
- Tokyo University of Science Unveils Synaptic Device for Low-Power Edge AI Sensors
- NetApp Joins the Vultr Cloud Alliance to Provide Scalable Data Management for Enterprise AI Workloads
- Anomalo Secures $10M to Advance Data Quality Solutions for Generative AI
November 22, 2024
- Telmai Partners with ZoomInfo to Enhance Data Quality Across Multi-Cloud Ecosystems
- BigID Redefines Privacy Operations with AI Automation
- HPE Advances Hybrid Cloud Leadership with Unified Virtualization Management, Fast Object Storage and Disconnected Private Clouds
- DataOps.live Achieves SOC 2 Type II Compliance
- LogicMonitor Gains $800M in Strategic Investment to Scale Global Operations