ANN - Wrong results returning as high as right ones!

[Yond]

We have been trying out the ANN index with both Multimodal and Text Gecko embeddings.

With Mulitmodal, the results are pretty poor - we took a bunch of pictures of dogs, buildings (cathedrals, houses) and other things

We then generated the image embedding using multimodal - created the index and tried searching the text for 'dog'. It didnt really work, some dogs came back and some houses and the scores were super low. Even with an image search it was not great

So we tried something else - we made longer text descriptions and tags of the images instead - described the dog in detail and the house and the image. 

Then we used Text Gecko and did the same process - made an index and tried searching.

This time the matches were higher, BUT using the term Dog - some houses (no dogs present) were brought back just as high and using Animal or Pet it was worse.

We are clearly doing something wrong - but not sure what... The examples we have seen work great!

Can anyone help to work this out please!

[nceniza]

Poor quality of result may be caused of number of things, from the quality of the inputs or may be the model endpoints etc. If possible, I would recommend Filing a case to have the opportunity to help improve the quality of results of the API :https://cloud.google.com/contact