How the vector database size calculator works
Each embedding is a list of numbers, so raw size is simply the count of vectors times dimensions times bytes per number. The search index then adds its own data structures.
Total = (Raw size × (1 + index overhead)) × Replicas
Common embedding sizes are 384, 768, 1,024, 1,536 and 3,072 dimensions. Graph indexes like HNSW store neighbour links for each vector, which typically adds roughly 10% to 100% depending on settings; metadata and the original text add more. Quantization (float16, int8 or binary) cuts memory by 2 to 32 times at some cost in recall. Results use decimal gigabytes (10⁹ bytes).
Worked example
One million 1,536-dimension float32 embeddings take 1,000,000 × 1,536 × 4 = 6.14 GB of raw data. With 50% index overhead you need about 9.22 GB of memory. Storing them as int8 instead would cut that to about 2.3 GB.
Frequently asked questions
How much memory does a vector database need?
Roughly vectors × dimensions × 4 bytes for float32, plus index overhead, metadata and headroom.
What is HNSW?
Hierarchical Navigable Small World, a graph index that gives fast approximate nearest-neighbour search at the cost of extra memory.
Does quantization hurt accuracy?
Slightly. int8 usually keeps recall close to float32; binary needs re-ranking to stay accurate.
Can vectors live on disk instead of RAM?
Yes. Disk-based indexes such as DiskANN keep most data on SSD with lower memory use.
Updated October 2026. Results are estimates for education, not financial or medical advice. All math runs in your browser; nothing you type is sent anywhere.