airflow.providers.common.ai.batch.dispatch¶

Dispatch from a connection and a model_id to a BatchAdapter.

The model_id prefix ("openai" in "openai:gpt-5") selects the adapter, because the prefix decides the request shape and the batch API being called. The connection type is then checked against the adapter’s conn_types, since the adapter has to know how to turn that connection’s fields into credentials. This is a separate decision from infer_model() (the synchronous Agent path), which has no notion of “does this provider have a batch API”.

Adapters are imported lazily, only once a request is dispatched to them, so importing this module never requires openai/anthropic to be installed. Another package can add an adapter in two ways: call register_adapter() at import time, or declare an entry point in the airflow.providers.common.ai.batch_adapters group whose name is the model prefix and whose value is module:Class.

Attributes¶

ENTRY_POINT_GROUP

Functions¶

`register_adapter`(adapter_cls)	Register an adapter class for its own `name` prefix.
`import_adapter_class`(name)	Return the adapter class for an already-resolved adapter name.
`split_model_id`(model_id)	Split `"<provider>:<model>"` into its two parts, rejecting a missing or empty half.
`get_adapter_class`(conn_type, model_id)	Resolve a connection type and `model_id` to an adapter class.
`resolve_adapter_name`(conn_type, model_id)	Return the adapter name for a connection/model pair. See `get_adapter_class()`.
`build_adapter_from_connection`(name, conn)	Instantiate an already-resolved adapter from a fetched connection.
`build_adapter`(name, *, llm_conn_id)	Fetch `llm_conn_id` and instantiate the adapter; the trigger's entry point, which only carries the id.

Module Contents¶

airflow.providers.common.ai.batch.dispatch.ENTRY_POINT_GROUP = 'airflow.providers.common.ai.batch_adapters'[source]¶

airflow.providers.common.ai.batch.dispatch.register_adapter(adapter_cls)[source]¶

Register an adapter class for its own name prefix.

Intended for other provider packages that ship a batch engine (Bedrock, Vertex, Azure OpenAI). A registration overrides a built-in or entry-point adapter with the same prefix.

airflow.providers.common.ai.batch.dispatch.import_adapter_class(name)[source]¶

Return the adapter class for an already-resolved adapter name.

airflow.providers.common.ai.batch.dispatch.split_model_id(model_id)[source]¶

Split "<provider>:<model>" into its two parts, rejecting a missing or empty half.

Raises:: UnsupportedBatchProviderError – model_id is None, has no ":", or has an empty prefix or model name.

airflow.providers.common.ai.batch.dispatch.get_adapter_class(conn_type, model_id)[source]¶

Resolve a connection type and model_id to an adapter class.

Raises:

UnsupportedBatchProviderError – no adapter serves the model_id prefix, or model_id is malformed.
BatchProviderNotYetSupportedError – an adapter exists but does not accept this connection type (e.g. pydanticai_azure today).

airflow.providers.common.ai.batch.dispatch.resolve_adapter_name(conn_type, model_id)[source]¶

Return the adapter name for a connection/model pair. See get_adapter_class().

airflow.providers.common.ai.batch.dispatch.build_adapter_from_connection(name, conn)[source]¶

Instantiate an already-resolved adapter from a fetched connection.

Uses the connection’s password/host fields as api_key/base_url, the same fields PydanticAIHook reads for the synchronous @task.llm path.

airflow.providers.common.ai.batch.dispatch.build_adapter(name, *, llm_conn_id)[source]¶

Fetch llm_conn_id and instantiate the adapter; the trigger’s entry point, which only carries the id.