Why it matters
Vector databases price by memory, and a million 1,536-dimension float32 vectors is already over 6 GB before any index links.
Quantisation and smaller embeddings cut that sharply; the table shows the trade-off.
How to use it
- Type or paste your input; the example shows the format.
- Press Estimate size. Results also update as you type.
- Read the summary line, then the table, notes or output below it.
- Copy or download what you need. Nothing leaves this page.
FAQ
How accurate is the HNSW estimate?
It counts 2×M neighbour links per vector at 4 bytes each, which is the bulk of the overhead. Real engines add some more.
Does int8 hurt recall?
Usually a little. Many engines rescore with full-precision vectors to recover it.
Is my input uploaded?
No. The tool runs in your browser and nothing you type, paste or open is sent anywhere.
Can I rely on the result?
It follows the public specification or formula named on this page and says where it simplifies. Check anything critical against the official tool or documentation before you ship it.