Work out vector database storage instantly with clear inputs, formula shown and shareable results.
Raw vector storage is simply count times dimensions times bytes per value, so five million 768-dimensional fp32 vectors need about 15 GB before any index exists. HNSW graphs add neighbour lists worth roughly 20 to 50 percent on top, and payload metadata adds more. Switching to fp16 or int8 quantisation halves or quarters the payload and is usually the first lever to pull.
Vector store size
raw = vectors x dimensions x bytes per value; total = raw x (1 + index overhead)
Moving fp32 to int8 cuts the payload to a quarter with a small recall loss; product quantisation can reach 10 to 30 times compression when recall can be traded for memory.
For low-latency HNSW search, effectively yes. Disk-based indexes such as DiskANN exist but add tens of milliseconds per query.