AI applications are pushing the limits of vector similarity search, requiring systems that can handle massive datasets while maintaining low latency and high throughput. ScyllaDB Vector Search addresses these needs by offering a unified platform for structured data and unstructured embeddings. Its architecture separates storage and indexing, allowing for ultra-low latency performance. In recent benchmarks with a dataset of 1 billion vectors, ScyllaDB achieved p99 latencies as low as 1.7 milliseconds and maintained a throughput of up to 252,000 queries per second under optimal conditions. The benchmarking involved two specific scenarios. The first aimed for ultra-low latency with moderate recall, suitable for applications like recommendation systems. Here, the system managed a p99 latency of 1.7 milliseconds while handling 30 concurrent searches. The second scenario focused on high recall for tasks requiring near-perfect accuracy, such as semantic search. In this case, ScyllaDB maintained a p99 latency below 12 milliseconds with a recall nearing 98%, while still delivering 6,500 queries per second. Integration of vector search within ScyllaDB reduces operational complexity. By storing metadata and embeddings together, it eliminates the need for separate systems, which often complicate data management and retrieval. This design allows users to perform complex queries that combine traditional search with semantic search. ScyllaDB's architecture ensures high availability and scalability while simplifying the overall data pipeline. The product is currently in General Availability, with plans for enhanced features like native filtering and memory optimizations in the coming months.