Upgrade complexity classification to 3-class (simple / medium / complex) #1
Description
Context
The current complexity classifier is a 2-class ML model (simple / medium).
"Complex" classification is handled via heuristic escalation.
This hybrid approach is intentional and documented, but leaves open the question
of whether ML-based "complex" prediction improves routing outcomes.
Open questions
- What features best distinguish "complex" queries from "medium"?
- Does ML-based complex prediction reduce unnecessary escalation?
- How does it affect cost savings vs quality?
Possible directions
- Collect 3-class labeled training data
- Retrain and evaluate a 3-class model
- Compare against the current hybrid ML + heuristic approach
This issue is exploratory and research-oriented.