Starling is an I/O-efficient disk-resident graph index framework designed for high-dimensional vector similarity search (HVSS) on data segments. The framework addresses the challenges of in-memory indexes, which struggle with the substantial increase in main memory requirements as vector data scales up. Starling optimizes data layout and search strategy within the segment, incorporating an in-memory navigation graph and a reordered disk-based graph with enhanced locality. It also employs a block search strategy to minimize costly disk I/O operations during vector query execution. Through extensive experiments, Starling demonstrates superior performance, achieving 43.9x higher throughput and 98% lower query latency compared to state-of-the-art methods while maintaining high accuracy. The framework is particularly effective in handling large-scale vector datasets on a single machine, accommodating tens of millions of vectors with 128 dimensions, and offering over 0.9 average precision and top-10 recall rates. Starling's key contributions include a novel data layout that improves vertex utilization ratio and reduces search path length, effective block shuffling algorithms to enhance data locality, and a block search strategy with computation-specific optimizations.Starling is an I/O-efficient disk-resident graph index framework designed for high-dimensional vector similarity search (HVSS) on data segments. The framework addresses the challenges of in-memory indexes, which struggle with the substantial increase in main memory requirements as vector data scales up. Starling optimizes data layout and search strategy within the segment, incorporating an in-memory navigation graph and a reordered disk-based graph with enhanced locality. It also employs a block search strategy to minimize costly disk I/O operations during vector query execution. Through extensive experiments, Starling demonstrates superior performance, achieving 43.9x higher throughput and 98% lower query latency compared to state-of-the-art methods while maintaining high accuracy. The framework is particularly effective in handling large-scale vector datasets on a single machine, accommodating tens of millions of vectors with 128 dimensions, and offering over 0.9 average precision and top-10 recall rates. Starling's key contributions include a novel data layout that improves vertex utilization ratio and reduces search path length, effective block shuffling algorithms to enhance data locality, and a block search strategy with computation-specific optimizations.