airflow.providers.common.ai.durable.caching_model¶

Caching model wrapper for durable execution.

Attributes¶

log

Classes¶

CachingModel

Wraps a model to cache responses in ObjectStorage for durable execution.

Module Contents¶

airflow.providers.common.ai.durable.caching_model.log[source]¶

class airflow.providers.common.ai.durable.caching_model.CachingModel(wrapped, *, storage, counter, replay_usage=None)[source]¶

Bases: pydantic_ai.models.wrapper.WrapperModel

Wraps a model to cache responses in ObjectStorage for durable execution.

On each request() call, checks if a cached response exists for the current step index and was produced by an equivalent request (same model, message history, settings, and tools – compared via fingerprint). If so, returns the cached response without calling the underlying model. Otherwise, calls the model and caches the response. A fingerprint mismatch means the agent changed between attempts; the stale entry is discarded and the step re-runs live.

With a replay_usage ledger, a replay hit does not count toward the run’s usage (see ReplayUsageLedger). The cached response is returned unchanged – it is part of what later steps’ fingerprints hash, so zeroing its usage would make every step after this one diverge and re-run live – and the ledger subtracts what the graph is about to add instead.

storage: airflow.providers.common.ai.durable.base.DurableStorageProtocol[source]¶

counter: airflow.providers.common.ai.durable.step_counter.DurableStepCounter[source]¶

replay_usage: airflow.providers.common.ai.durable.replay_usage.ReplayUsageLedger | None = None[source]¶

credit_first_replay()[source]¶

Credit the run’s first model request if the cache holds a response for it.

Called before the run starts, because pydantic-ai checks request_limit before the first request reaches this model; without the credit, a retry whose seeded requests already equals request_limit could not start even when every step would replay for free.

async request(messages, model_settings, model_request_parameters)[source]¶

Make a request to the model.

This is ultimately called by pydantic_ai._agent_graph.ModelRequestNode._make_request(…).