Just want to say a huge thanks for the AI image-search demo. Been wanting to build something like this for a while - since seeing the KNN query support.
… just haven’t managed to figure out all the bits ot make it work, so seeing a running demo, is really useful to piece it altogether.
Have been able to get the demo running relatively easily. Imported about 250,000 images from Geograph Britain and Ireland.
… which is landscape photography from the British Isles, so is a very good fit for searching by image.
For now there is an public instance available here: https://image-test.geograph.org.uk/
… this seems to be running easily in a t2.small instance on AWS.
… manticore seems to be using just short of 500Mb of memory, for hosting the 250k image embedding index.
Its a good dataset to test the ‘a sunset over the beach’ query 
as it has lots of images of sunsets and beaches. Other queries that work well are things like ‘busy street’ or, ‘level crossing’ - all matched on the image embeddings, rather than the textual metadata of the image.
Have put my changes on a fork on github
in case it helps others. Didnt need to do much to get the demo running, and inlcudes an adaptation of the load-dataset script, as it didnt run out of the box, needed to add some ‘includes’, but otherwise it runs just fine within the ‘php’ container.