Charles Xie is the founder and CEO of Zilliz, focusing on building next-generation databases and search technologies for AI and LLMs applications. At Zilliz, he also invented Milvus, the world’s most popular open-source vector database for production-ready AI. He is currently a board member of LF AI & Data Foundation and served as the board’s chairperson in 2020 and 2021. Charles previously worked at Oracle as a founding engineer of the Oracle 12c cloud database project. Charles holds a master’s degree in computer science from the University of Wisconsin-Madison.
Zilliz is the team behind LF AI Milvus®, a widely used open-source vector database. The company focuses on simplifying data infrastructure management, aiming to make AI more accessible to corporations, organizations, and individuals alike.
Can you share the story behind founding Zilliz and what inspired you to develop Milvus and focus on vector databases?
My journey in the database field spans over 15 years, including six years as a software engineer at Oracle, where I was a founding member of the Oracle 12c Multitenant Database team. During this time, I noticed a key limitation: while structured data was well-managed, unstructured data—representing 90% of all data—remained largely untapped, with only 1% analyzed meaningfully.
In 2017, the growing ability of AI to process unstructured data marked a turning point. Advances in NLP showed how unstructured data could be transformed into vector embeddings, unlocking its semantic meaning. This inspired me to found Zilliz, with a vision to manage “zillions of data.” Vector embeddings became the cornerstone for bridging the gap between unstructured data and actionable insights. We developed Milvus as a purpose-built vector database to bring this vision to life.
Over the past two years, the industry has validated this approach, recognizing vector databases as foundational for managing unstructured data. For us, it’s about more than technology—it’s about empowering humanity to harness the potential of unstructured data in the AI era.
How has the journey of Zilliz evolved since its inception six years ago, and what key challenges did you face while pioneering the vector database space?
The journey has been transformative. When we started Zilliz seven years ago, the real challenge wasn’t fundraising or hiring—it was building a product in completely uncharted territory. With no existing roadmaps, best practices, or established user expectations, we had to chart our own course.
Our breakthrough came with the open-sourcing of Milvus. By lowering barriers to adoption and fostering community engagement, we gained invaluable user feedback to iterate and improve the product. When Milvus launched in 2019, we had around 30 users by year-end. This grew to over 200 by 2020 and nearly 1,000 soon after.
Today, vector databases have shifted from a novel concept to essential infrastructure in the AI era, validating the vision we started with.
As a vector database company, what unique technical capabilities does Zilliz offer to support multimodal vector search in modern AI applications?
Zilliz has developed advanced technical capabilities to support multimodal vector search:
- Hybrid Search: We enable simultaneous searches across different modalities, such as combining an image’s visual features with its text description.
- Optimized Algorithms: Proprietary quantization techniques balance recall accuracy and memory efficiency for cross-modal searches.
- Real-Time and Offline Processing: Our dual-track system supports low-latency real-time writes and high-throughput offline imports, ensuring data freshness.
- Cost Efficiency: Our Extended Capacity instances leverage intelligent Tiered Storage to reduce storage costs significantly while maintaining high performance.
- Embedded AI Models: By integrating multimodal embedding and ranking models, we’ve lowered the barrier to implementing complex search applications.
These capabilities allow developers to efficiently handle diverse data types, making modern AI applications more robust and versatile.
How do you see Multimodal RAG advancing AI’s ability to handle complex real-world data like images, audio, and videos alongside text?
Multimodal RAG (Retrieval-Augmented Generation) represents a pivotal evolution in AI. While text-based RAG has been prominent, most enterprise data spans images, videos, and audio. The ability to integrate these diverse formats into AI workflows is critical.
This shift is timely, as the AI community debates the limits of available internet text data for training. While text data is finite, multimodal data remains vastly underutilized—ranging from corporate videos to Hollywood films and audio recordings.
Multimodal RAG unlocks this untapped reservoir, enabling AI systems to process and leverage these rich data types. It’s not just about addressing data scarcity; it’s about expanding the boundaries of AI’s capabilities to better understand and interact with the real world.
How does Zilliz differentiate itself from competitors in the rapidly growing vector database market?
Zilliz stands out through several unique aspects:
- Dual Identity: We are both an AI company and a database company, pushing the boundaries of data management and AI integration.
- Cloud-Native Design: Milvus 2.0 was the first distributed vector database to adopt a disaggregated storage and compute architecture, enabling scalability and cost-efficiency for over 100 billion vectors.
- Proprietary Enhancements: Our Cardinal engine achieves 3x the performance of open-source Milvus and 10x over competitors. We also offer disk-based indexing and intelligent Tier Storage for cost-effective scaling.
- Continuous Innovation: From hybrid search capabilities to migration tools like VTS, we’re constantly advancing vector database technology.
Our commitment to open source ensures flexibility, while our managed service, Zilliz Cloud, delivers enterprise-grade performance with minimal operational complexity.
Can you elaborate on the significance of Zilliz Cloud and its role in democratizing AI and making vector search services accessible to small developers and enterprises alike?
Vector search has been used by tech giants since 2015, but proprietary implementations limited its broader adoption. At Zilliz, we’re democratizing this technology through two complementary approaches:
- Open Source: Milvus allows developers to build and own their vector search infrastructure, lowering technical barriers.
- Managed Service: Zilliz Cloud eliminates operational overhead, offering a simple, cost-effective solution for businesses to adopt vector search without requiring specialized engineers.
This dual approach makes vector search accessible to both developers and enterprises, enabling them to focus on building innovative AI applications.
With advancements in LLMs and foundation models, what do you believe will be the next big shift in AI data infrastructure?
The next big shift will be the wholesale transformation of AI data infrastructure to handle unstructured data, which makes up 90% of the world’s data. Existing systems, designed for structured data, are ill-equipped for this shift.
This transformation will impact every layer of the data stack, from foundational databases to security protocols and observability systems. It’s not about incremental upgrades—it’s about creating new paradigms tailored to the complexities of unstructured data.
This transformation will touch every aspect of the data stack:
- Foundational database systems
- Data pipelines and ETL processes
- Data cleaning and transformation mechanisms
- Security and encryption protocols
- Compliance and governance frameworks
- Data observability systems
We’re not just talking about upgrading existing systems – we’re looking at building entirely new paradigms. It’s like moving from a world optimized for organizing books in a library to one that needs to manage, understand, and process the entire internet. This shift represents a total new world, where every component of data infrastructure might need to be reimagined from the ground up.
This revolution will redefine how we store, manage, and process data, unlocking vast opportunities for AI innovation.
How has the integration of NVIDIA GPUs influenced the performance and scalability of your vector search?
The integration of NVIDIA GPUs has significantly enhanced our vector search performance in two key areas.
First, in index building, which is one of the most compute-intensive operations in vector databases. Compared to traditional database indexing, vector index construction requires several orders of magnitude more computational power. By leveraging GPU acceleration, we’ve dramatically reduced index-building time, enabling faster data ingestion and improved data visibility.
Second, GPUs have been crucial for high-throughput query use cases. In applications like e-commerce, where systems need to handle thousands or even tens of thousands of queries per second (QPS), GPU’s parallel processing capabilities have proven invaluable. By utilizing GPU acceleration, we can efficiently process these high-volume vector similarity searches while maintaining low latency.
Since 2021, we’ve been collaborating with NVIDIA to optimize our algorithms for GPU architecture, while also developing our system to support heterogeneous computing across different processor architectures. This gives our customers the flexibility to choose the most suitable hardware infrastructure for their specific needs.
As vector databases play a critical role in AI, do you see their application extending beyond traditional use cases like recommendation systems and search to industries like healthcare?
Vector databases are rapidly expanding beyond traditional applications like recommendation systems and search, penetrating industries we never imagined before. Let me share some examples.
In healthcare and pharmaceutical research, vector databases are revolutionizing drug discovery. Molecules can be vectorized based on their functional properties, and using advanced features like range search, researchers can discover all potential drug candidates that might treat specific diseases or symptoms. Unlike traditional top-k searches, range search identifies all molecules within a certain distance of the target, providing a comprehensive view of potential candidates.
In autonomous driving, vector databases are enhancing vehicle safety and performance. One interesting application is in handling edge cases – when unusual scenarios are encountered, the system can quickly search through massive databases of similar situations to find relevant training data for fine-tuning the autonomous driving models.
We’re also seeing innovative applications in financial services for fraud detection, cybersecurity for threat detection, and targeted advertising for improved customer engagement. For instance, in banking, transactions can be vectorized and compared against historical patterns to identify potential fraudulent activities.
The power of vector databases lies in their ability to understand and process similarity in any domain – whether it’s molecular structures, driving scenarios, financial patterns, or security threats. As AI continues to evolve, we’re just scratching the surface of what’s possible. The ability to efficiently process and find patterns in vast amounts of unstructured data opens up possibilities we’re only beginning to explore.
How can developers and enterprises best engage with Zilliz and Milvus to leverage vector database technology in their AI projects?
There are two main paths to leverage vector database technology with Zilliz and Milvus, each suited for different needs and priorities. If you value flexibility and customization, Milvus, our open-source solution, is your best choice. With Milvus, you can:
- Experiment freely and learn the technology at your own pace
- Customize the solution to your specific requirements
- Contribute to development and modify the codebase
- Maintain complete control over your infrastructure
However, if you want to focus on building your application without managing infrastructure, Zilliz Cloud is the optimal choice. It offers:
- An out-of-the-box solution with one-click deployment
- Enterprise-grade security and compliance
- High availability and stability
- Optimized performance without operational overhead
Think of it this way: if you enjoy ‘tinkering’ and want maximum flexibility, go with Milvus. If you want to minimize operational complexity and get straight to building your application, choose Zilliz Cloud.
Both paths will get you to your destination – it’s just a matter of how much of the journey you want to control versus how quickly you need to arrive
Thank you for the great interview, readers who wish to learn more should visit Zilliz or Milvus.