How Poizon Accelerates AI-Powered Shopping with Vector Search at Billion Scale Using Zilliz Cloud
Sub-90ms response times
for billion-scale vector searches
Lower total cost
than self-managed setup
3-cluster compexity eliminated
for production
Innovation-focused
instead of maintainance
About Poizon
Poizon is one of Asia’s fastest-growing social e-commerce platforms, evolving from a sneaker-focused marketplace into a comprehensive fashion and lifestyle destination. Serving millions of users daily, Poizon leverages AI-powered capabilities—including visual search, automated authenticity verification, personalized recommendations, and fraud prevention—to deliver seamless and trusted shopping experiences.
Behind the scenes, Poizon processes massive volumes of visual and textual data to power marketplace operations, product authentication, and community features. Delivering instant and accurate product discovery is central to its competitive edge—making high-performance vector search a critical foundation for customer satisfaction and business growth..
The Challenge: Building Infrastructure for Billion-Scale AI Workloads
Poizon relies on GenAI technologies across multiple mission-critical scenarios:
Image Search – allowing users to instantly find similar products through image uploads
AI-Powered Authentication – automatically verifying the authenticity of luxury goods
Algorithm Optimization – strengthening recommendation engines and product discovery
Security and Risk Control – preventing fraud and protecting the platform
Each of these capabilities depends on a robust vector database infrastructure. Poizon’s leadership recognized that its long-term competitiveness would hinge on delivering fast, intelligent, and trustworthy AI-driven experiences. To achieve this goal, the team needed an infrastructure capable of handling billion-scale vector operations while maintaining sub-90ms response times—performance levels essential for user engagement.
A Hybrid Vector Database Strategy for Performance and Cost Efficiency
Poizon tested various options, including Milvus and Qdrant with real data, replaying typical queries and measuring two latency targets: sub-90 ms for mission-critical paths and sub-500 ms for very large datasets. They also compared index choices—HNSW as the default, DiskANN for large/low-QPS sets—plus Kubernetes scaling, day-two operations, and community support.
Milvus was the better fit. It matched the team’s Go-based stack, scaled cleanly on Kubernetes, and had an active domestic community that sped up troubleshooting and knowledge sharing. It performed reliably in soak tests, and the team could adapt tools when needed—for example, adjusting milvus-backup to skip problematic segments. Milvus became the standard across Poizon’s vector workloads.
As scale grew, Poizon revisited how to hit ultra-low latency. Benchmarks, traffic shadowing, and A/B tests showed a clear split: self-managed Milvus with DiskANN is ideal for big datasets with moderate latency goals, while the managed Milvus (also known as Zilliz Cloud) consistently meets the sub-90 ms bar at billion scale with less operational effort.
In Poizon’s environment, pushing self-managed clusters further led to diminishing returns—latency leveled off around ~200 ms even after scaling to 60 QueryNodes, largely due to QueryNode–Proxy interactions. Achieving sub-90 ms on-prem would have required at least three parallel clusters and complex multi-read/multi-write logic—possible, but expensive and heavy to run.
Milvus proved the best fit—matching Poizon’s Go-based stack, scaling smoothly on Kubernetes, and backed by an active community. Poizon adopted a hybrid approach: Zilliz Cloud for latency-critical workloads (such as visual search and AI authentication) and self-managed Milvus for cost-efficient scenarios with steady sub-500 ms performance—delivering both cloud-scale speed and operational efficiency.
The Results: Faster Search, Leaner Architecture, Greater Innovation
The adoption of Zilliz Cloud delivered immediate, measurable improvements across Poizon’s AI-powered commerce platform:
Sub-90ms Performance – Zilliz Cloud consistently achieved the ultra-low latency required for visual search and AI authentication—something self-managed Milvus couldn’t deliver, even after scaling to 60 QueryNodes, where performance plateaued at ~200ms.
Simpler, More Cost-Effective Architecture – By eliminating the need for a complex three-cluster setup and multi-write/multi-read logic, Zilliz Cloud not only simplified operations but also reduced costs. The managed service outperformed what a costly self-managed alternative could have achieved.
Engineering Focus Redirected to Innovation – With infrastructure bottlenecks resolved, Poizon’s engineering team shifted away from database tuning and maintenance, and instead focused on building the AI-powered features that strengthen their competitive edge in e-commerce.
Looking Forward: Building Data Pipeline Excellence
With a high-performance vector database foundation in place, Poizon is now focused on building a best-in-class data pipeline to fuel the next stage of AI innovation.
The engineering team is creating automated migration and ingestion tools that will allow business teams to focus on data-driven applications while DBAs manage quantization and ingestion behind the scenes. This will accelerate the rollout of new AI features across business units without the burden of manual preparation.
In parallel, Poizon is developing data consistency validation tools—together with internal teams and the Milvus community—to ensure alignment between upstream systems like MySQL and their vector databases. These efforts will safeguard data integrity across the AI pipeline, laying the foundation for scalable, trustworthy AI innovation.
Conclusion
Poizon’s journey illustrates how the right vector database strategy can unlock AI innovation at scale. By combining Zilliz Cloud for performance-critical workloads with self-managed Milvus for cost-optimized scenarios, the company removed infrastructure bottlenecks and freed its engineers to focus on delivering differentiated AI features.
For e-commerce companies pursuing AI-powered experiences, Poizon’s hybrid approach demonstrates that infrastructure is not just a technical choice—it is a competitive advantage. With performance and reliability assured, the path to breakthrough user experiences becomes inevitable.
- About Poizon
- The Challenge: Building Infrastructure for Billion-Scale AI Workloads
- A Hybrid Vector Database Strategy for Performance and Cost Efficiency
- The Results: Faster Search, Leaner Architecture, Greater Innovation
- Looking Forward: Building Data Pipeline Excellence
- Conclusion
Content
Use case
Industry
E-commerce