Vector storage calculator

What it does

It works out the bytes a set of vectors takes: the vectors themselves at the precision you choose, the metadata stored beside each one, an index overhead you enter, and the number of copies you keep. It shows the same set at every precision, so the saving from quantising is in front of you.

Try it: 10 million vectors of 1,536 dimensions, 1 million chunks stored with their text, or 100 million binary vectors, replicated.

The formulas

• One vector: dimensions × 4 bytes as float32, × 2 as float16 or bfloat16, × 1 as int8, or ⌈dimensions ÷ 8⌉ as binary — one bit a dimension, rounded up to whole bytes.
• One copy: vectors × (one vector + metadata per vector) × (1 + overhead ÷ 100).
• Total: one copy × copies.

Sizes are given in decimal units (1 GB = 10⁹ bytes), as disks and most bills count, and in binary units (1 GiB = 2³⁰ bytes), as memory is counted.

What it leaves to you

• The index overhead. A graph index, an inverted-file index and a flat index each add a different amount, which depends on its settings — the number of links per node, the number of lists — and on the database. The tool does not guess it for any product: enter the figure your database documents, or 0 for the raw size.
• Metadata. Whatever you store with each vector: its id, the chunk's text, its fields. Text is roughly one byte a character in English, more in scripts outside ASCII.
• Precision loss. int8 and binary vectors take less space and give up some accuracy; how much depends on the model and the data, and is worth measuring with your own queries.

What it does not do

It is arithmetic, not a price: it names no database and no bill. Salyro does not serve embeddings or store vectors, and nothing you enter is sent anywhere.