I built this same solution for myself last year, used Hugging Face's "SmolVLM". It works surprisingly well. I use the model to generate verbose descriptions of each image, embed the descriptions using another model, which I also use for the query embedding.
The stack is hacky, since it was mostly for myself...