Query Routing on enterprise developer-support dataset (test)
89.83Overall AccuracyFlyRoute (flywheel)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FlyRoute (flywheel)Backbone LLM=Qwen3-8B, Retrieval Method=BM25 + LLM-fewshot, Training/Evolution Phase (Cold vs Flywheel)=Flywheel (after all training queries)2026.05 | 89.83 | 94.26 | 88.7 | 81.29 | 91.18 | |
| FlyRoute (cold)Backbone LLM=Qwen3-8B, Retrieval Method=BM25 + LLM-fewshot, Training/Evolution Phase (Cold vs Flywheel)=Cold (seed-only stores)2026.05 | 78.04 | 82.3 | 78.42 | 69.05 | 77.45 | |
| LLM RouterBackbone LLM=Qwen3-8B, Retrieval Method=None (Static seed descriptions only), Training/Evolution Phase (Cold vs Flywheel)=Zero-shot2026.05 | 72.57 | 93.61 | 62.33 | 37.41 | 77.45 |