Type a description, CLIP ranks the gallery by meaning — not keywords or filenames. The model (openai/clip-vit-base-patch32) downloads once and runs entirely in your browser from then on.
This isn't a precomputed lookup table. @huggingface/transformers (transformers.js) loads a
real ONNX export of CLIP and runs it via WebGPU/WASM directly in your browser tab. For each search, every
image in the gallery is scored against your exact query text through
pipeline('zero-shot-image-classification', ...) — change the query, the rankings change,
computed fresh, client-side, no server round-trip. The Python reference implementation in
python/clip_search.py runs the same underlying model offline using
transformers.CLIPModel.