Nanodb r36.4.0 Search queries return an incorrect list of images

Nanodb docker image: dustynv/nanodb:r36.4.0 gives wrong results for search queries, also gives incorrect or duplicate index warning when performing validation.

@dusty_nv

22:15:22 | WARNING | incorrect or duplicate index [9999] indexes=[-1 -1 -1 -1 -1 -1 -1 -1] distances=[-3.4028235e+38 -3.4028235e+38 -3.4028235e+38 -3.4028235e+38
-3.4028235e+38 -3.4028235e+38 -3.4028235e+38 -3.4028235e+38]

https://hub.docker.com/layers/dustynv/nanodb/r36.4.0/images/sha256-1a8ca9271e8da9a0eb19fa1801b63a2563737d0928016f614e403f88e813fc1a

result:

Hi,

Could you check the lists to see if all the images are incorrect or just #570739?

Thanks.

all of the results are incorrect.
I ran the nanodb code from dustynv/nanodb:r36.2.0 in dustynv/nanodb:r36.4.0
it seems like the vectors.bin is not generated properly.
it works but is slower.
The implementation of nanodb in both images is different.

Hi,

Do you mean pulling the dustynv/nanodb:r36.2.0 instead of dustynv/nanodb:r36.4.0?

One of our users claims the r36.2.0 is faster.

Not sure if this is data-dependent.

Does using the r36.2.0 container work for you?
Thanks.

I mean, copying the code for nanodb from dustynv/nanodb:r36.2.0 and running it in on dustynv/nanodb:r36.4.0 , I noticed it uses a different omplementation of CLIPEmbedding in nanodb.

Hi,

Thanks for this info.

We will check with the owner and get back to you.
But the code from r36.2.0 already works for you, is that correct?

Thanks.

Yes code from r36.2.0 gives the correct results when searching