Vector image search

Find visually similar trading cards using a precomputed image embedding.

Use this endpoint for:

• image-based card recognition flows
• camera upload search
• similarity search from a cropped card artwork image

Preparation requirements:

• Generate a single image embedding from one card image at a time.
• Use CLIP ViT-B/32-compatible image preprocessing so the model input is prepared as float32 pixel values with shape [1, 3, 224, 224].
• Use the model embedding output with shape [1, 512].
• L2-normalize the 512-dimensional vector before sending it to this API.
• Send only the normalized numeric vector in vectors. Do not send raw pixels, base64 image data, or wrapped model response objects.

Behavior:

• Accepts exactly one 512-dimensional embedding per request.
• Accepts an optional limit between 1 and 10. If omitted, the API returns up to 5 matches.
• Returns visually similar card matches, ordered by similarity score.
• Each match includes the scored card identifier fields needed to resolve the card and display a primary image preview.

Good to know:

• This route is designed for image embeddings only. It does not generate embeddings from uploaded files.
• Invalid vectors are rejected before any similarity search is performed.

{
  "limit": 3
}