Classification and routing
One label per input, at the volume your queue actually has.
Routing tickets, detecting intent and triaging requests are narrow, repetitive decisions. A small model trained on your own labelled history learns your categories exactly, instead of approximating them from a long prompt on every call.
The problem
Every request pays for a large model
A label is a few tokens, but each call still sends the whole prompt and input to a general-purpose model.
Categories drift from the prompt
Edge cases between similar labels are where general models disagree with your team. Your history already contains the right answers.
Latency sits in the critical path
Routing happens before anything else, so a slow call delays everything after it.
Where it applies
- Support ticket routing
- Intent detection for assistants and IVR
- Triage and priority
- Content moderation categories
- Lead and email categorisation
How the switch is checked
Before you switch, the fine-tuned model and the LLM you use today answer the same held-out examples from your data.
- Accuracy and macro-F1 on held-out examples, for the fine-tuned model and the LLM you use today
- Accuracy for every label, so a weak category is visible before you switch
- Answers that don't name exactly one of your labels are counted, not hidden
What you need
A file with the input text and its label for each example, e.g. instruction and output columns.