Dense vector disk size

[amberry]

Hi there,

We're attempting to index about 10,000,000 1536-dim vectors. Per the documentation, we've calculated this should take up 10M*(4*1536+4) = approximately 61.5GB. In reality, this index is taking up about 250GB. (There are several other keyword/text fields, a mix of indexed/unindexed, but we don't believe these will be the reason the size is exploding?)

What are we doing wrong/where have we gone wrong in our calculations? Alternatively, is there anything we can do to reduce this disk size?

We also tried to set index_type to int4_hnsw , which unexpectedly led to a larger rather than smaller size.
For reference, we are on version 8.15.0.

[Kathleen_DeRusso]

Welcome @amberry !

Off the cuff, your vector math seems reasonable but it's hard to say how much impact your other fields would have.

Here's what I would check:

• Do you have a lot of unmerged segments with deleted documents that could be inflating your disk size?
• You can see if anything glaring comes out of the Disk usage API
• You could look at some other suggestions for tuning kNN search or just our general recommendations for tuning for disk usage