Skip to main content
Documentation

Experience

Overview

Docs indexEdit page
Version: Latest

Experience

Overview​

Experience is the online Router Learning evidence used by adaptation. It is kept in process on the request path and summarized from bounded outcomes and telemetry. Offline recipe learning can export seed-pack artifacts for cold-start analysis and future warmup flows, but the first public API does not expose a runtime seed-pack import switch.

Key Advantages​

  • Keeps expensive aggregation out of the request path.
  • Turns outcomes and telemetry into compact read-time evidence.
  • Gives adaptation a typed place for learned model-choice facts.
  • Preserves Router Replay as the event log source of truth.

What Problem Does It Solve?​

Some routing evidence needs past requests: model fit, overuse, provider failures, latency, cache reuse, and effective cost. Reading all events during a request would be too slow. The router updates compact model experience in memory and writes durable evidence to Router Replay for offline analysis.

When to Use​

  • Adaptation needs aggregate evidence from outcomes and runtime telemetry.
  • An operator wants offline evals to explain cold-start model quality evidence.
  • Offline recipe learning needs seed-pack artifacts for experiments or future warmup flows.

Configuration​

The current public API does not expose experience.enabled, experience.source, or a runtime seed-pack import field. If adaptation is enabled, model experience is part of the routing_sampling implementation.

Online experience is keyed by matched decision, decision tier, and model:

decision_id + decision_tier + model
  -> decision_tier + model
  -> model

Outcome ingestion updates model-targeted experience. Route, policy, stability, provider, and router outcomes feed replay diagnostics and offline recipe learning instead of directly mutating model quality.