Heuristic Router
The heuristic router uses a rule-based scoring system to select the optimal model for each query. It analyzes query features (length, complexity, tool requirements) and matches them against model capability profiles.
Fonctionnement
- Feature extraction -- analyze the query for length, keyword patterns, and tool requirements
- Capability scoring -- score each model's suitability based on extracted features
- Cost weighting -- apply cost preferences to favor cheaper models for simple queries
- Selection -- choose the highest-scoring model that meets minimum quality thresholds
Scoring Factors
| Factor | Weight | Description |
|---|---|---|
| Capability match | 0.4 | Model's ability to handle the query type |
| Elo rating | 0.2 | Historical model performance |
| Cost efficiency | 0.2 | Tokens per dollar |
| Latency | 0.1 | Average response time |
| Context window | 0.1 | Fits the conversation context |
Configuration
toml
[router]
strategy = "heuristic"
[router.heuristic]
complexity_threshold = 0.6
prefer_cheap_below = 0.4
elo_weight = 0.2
cost_weight = 0.2