Tip: working effectively with models

Use AdaptiveModelRouter.get_best_model() to select a model rather than hardcoding model IDs — it ranks candidates by quality_score (derived from success_rate, avg_latency_ms, and avg_cost) against your actual telemetry, so routing improves automatically as your workflows accumulate history.

Why: Hardcoded model IDs bypass the circuit breaker and fallback chain, meaning a single provider outage can halt a workflow that AdaptiveModelRouter would have recovered from silently.

Tradeoff: get_best_model() requires a min_success_rate threshold (default 0.8) and a populated telemetry store. If your sample_size is too low, the router falls back to registry defaults anyway — so the benefit only compounds after enough runs to produce reliable ModelPerformance data.

Tags: models, auth, llm