How does it interpret the images

right now it's actually just OCR on top of images, so only images with text.

I'm actually seeing if we can do a deeper integration with embeddings though. Googles new Vertex api has a text+image embedding model which may work nicely