I like RAG for policy docs and vendor onboarding PDFs. I do not like it as a costume on a SKU matcher. Embeddings of text about an item are not embeddings of the item.
If the answer lives in a row, query the row. If the answer lives in a photo, match the photo. Language on top can explain the row after you fetched it.
Hallucinated SKUs are a special hell
A model that invents a code you almost have is worse than unknown. It looks plausible in a font you trust.
Use RAG where the corpus is actually language
SOPs, return policies, "how do I process a hazmat exception." Point it at documents with owners. Keep it away from on-hand counts.
