Expand description
Classifiers for routing decisions.
Two-stage pipeline:
- Heuristic — fast, no LLM call, based on language-agnostic structural signals.
- LLM (optional) — refines ambiguous cases by asking a small model.
The heuristic stage uses only quantitative/structural signals: message length, line count, code blocks, file paths, symbol density, question form, context tokens, turn count. No keyword matching.
The LLM stage is only activated when classifier_model is configured
and the heuristic score falls in the ambiguous zone (0.25–0.75).
Structs§
- Classifier
Input - Input for the classifier — metadata about the current request.
- Heuristic
Classifier - Fast classifier based entirely on language-agnostic structural signals.
- LlmClassifier
- LLM-based classifier for refining ambiguous heuristic scores.