What is visual search?
Visual search is search where the query is a picture. A user photographs an object, and the system returns the catalog items or documents that look most like it. A vision model converts every image in the collection to an embedding, a list of numbers that places visually similar images near each other. The query photo is converted the same way, and a vector search returns its nearest neighbors, even when nobody has tagged the item.
Image recognition assigns a label from a fixed list, while visual search ranks similarity against whatever is in the index, so new items are searchable the moment they are added, with no retraining.
Visual search in practice
An electrical supply distributor fields hundreds of calls a day from contractors who need a replacement part but have no part number. With visual search in the mobile app, the contractor photographs the old breaker or fitting and receives the closest catalog matches plus compatible brands. Counter calls drop and order accuracy improves. A furniture retailer uses the same technique to let shoppers find a sofa similar to one in a magazine photo.
What to watch for
- Catalog image quality sets the ceiling; one studio photo per item will match poorly against a photo taken in a dim electrical room.
- Decide whether similarity means the same item, the same style, or a compatible item, and test the ranking against that definition.
- Embedding models can run on your own infrastructure, which matters when the images are customer uploads or internal records.