Put everything where it belongs.

Your categories, not a fixed taxonomy — so the output drops straight into the structure your system already uses, with no mapping layer in between.

Given an incoming item, which of your categories does it belong to?

Text, images, or rows of structured fields.

What makes this hard by hand.

Manual filing doesn't keep up
A steady stream of new items means a steady stream of someone deciding where each one goes.
Generic taxonomies don't fit
Off-the-shelf tagging returns labels that don't match your database. You end up mapping one vocabulary onto another and losing precision.
Bad organization compounds
Miscategorized items are hard to find later, and the cost shows up in search, reporting, and recommendations.

Three steps.

1

Define your categories

Use the exact labels your system already uses. No mapping layer, no translation.

2

Start before you have training data

The function answers from the first call using the label names alone, so you can wire it up and see how it behaves on real items immediately.

3

Correct as you go

Fix the ones it gets wrong. Those corrections train a private model on your taxonomy.