Skip to content

learned_router

diapason.learning.routing.learned_router

Learned router policy — trace-driven query_class -> model mapping.

Merges the runtime select_model() logic from TraceDrivenPolicy with the batch update() logic from SFTRouterPolicy into a single class registered as "learned" in RouterPolicyRegistry.

LearnedRouterPolicy

Bases: RouterPolicy

Trace-driven router that learns query_class -> model mappings.

Implements RouterPolicy.select_model() for runtime routing AND provides update_from_traces() / observe() for learning from traces, plus update() for batch learning from a trace store.

policy_map property

Current learned routing decisions (read-only copy).

observe(query, model, outcome, feedback)

Record a single observation for online (incremental) updates.

select_model(context)

Select the best model based on learned policy or fallback.

update(trace_store, **kwargs)

Batch update: analyze trace outcomes and update the policy map.

This is the batch learning interface (from the old SFTRouterPolicy).

update_from_traces(*, since=None, until=None)

Recompute the policy map from trace history via TraceAnalyzer.

ensure_registered()

Register LearnedRouterPolicy if not already present.