Calculate the storage size required for a set of vector embeddings.
Vector storage size is simply the number of vectors multiplied by dimensions and bytes per element. One million 1536-dimensional FP32 vectors requires about 5.7 GB of raw storage. Actual database size includes index overhead (typically 20-100% extra).
Vector Storage
size = num_vectors × dimensions × bytes_per_float
Yes — ANN indexes (HNSW, IVF) typically add 20-100% overhead on top of raw vector storage for graph structures, quantization codebooks, and metadata.
Yes — product quantization (PQ) or scalar quantization can reduce storage by 4-8× with minimal recall loss. Binary quantization gives 32× reduction but significantly impacts accuracy.