Used storage space of binary quantized vector

Support
,
1

Hi, I’m thinking about reducing the needed space using BQ, but the documentation didn’t tell if the original vector is saved and additionally the BQ version in the index or if only the space used by the BQ version is counted for your price calculation in Weaviate Cloud.

2

Hi @Alexander_Kordecki !!

Welcome to our community :hugs:

Yes, both the original and the quantized vectors are stored in disk. So while quantization/compression will save on memory, it will increase in disk usage and slightly decrease in accuracy, as tradeoffs.

In our serverless offer, we will only charge for the stored vectors in memory.

This means that if you have a vector with 1536 dimensions, it will be charge accordingly.

And if you compress that, using a quantization, let’s say by a factor of 8, to 192 dimensions, it will be charged the 192 dimensions.

Let me know if this helps!

Thanks!

3

Hi Duda,

since i’m not completely sure, I understand you correctly, since quantization is not reducing the dimensionality:
When I have a 1536 dimensions vector, with FP32, which makes 6KB per vector, I have to pay for 1536 dimensions.
When I have a 1536 dimensions vector with BQ, which makes 192B per vector, I have to pay for 1536 dimensions - the same price. Correct?

Thanks,
Alex

4

Hi @Alexander_Kordecki !!

Sorry, I missed the big BQ part entirely!

Indeed, BQ will not reduce the dimensions. So billing for serverless will be the same in this scenario.

For our enterprise offering, as we the cost is not calculated only on stored dimensions, it will have an impact on billing.

Let me know if this helps!

thanks!