Giant models are slow, expensive, and eager to bluff. A small visual embedder plus nearest neighbors has been enough for SKU lookup when the library is honest.
I can explain nearest neighbors to a manager. I cannot honestly explain why a chat model invented a bin.
Cost is an operations constraint
If every scan costs a cloud inference that adds up at volume, someone will turn the tool off. Local or cheap embeddings keep the lights on.
Use the big models at the edge of the map
Language is useful for messy vendor emails and for drafting exception notes. Identity of a physical item is a retrieval problem. Mix them with intent, not because a slide said AI.
