Point at a card. Find the match.

The server fetches the image, embeds it through SigLIP-Base, and asks pgvector for the closest matches. On mobile this happens on-device with MobileCLIP and the float[] travels instead of the bytes.