Consistently use llm ask methods

guardrails/async_guard.py

@@ -325,7 +325,14 @@ async def _exec(
         Returns:
             The raw text output from the LLM and the validated output.
         """
-        api = get_async_llm_ask(llm_api, *args, **kwargs)  # type: ignore
+        api = None
+
+        if llm_api is not None or kwargs.get("model") is not None:
+            api = get_async_llm_ask(llm_api, *args, **kwargs)  # type: ignore
+
+        if self._output_formatter is not None:
+            api = self._output_formatter.wrap_async_callable(api)  # type: ignore
+
         if kwargs.get("stream", False):
             runner = AsyncStreamRunner(
                 output_type=self._output_type,

guardrails/formatters/base_formatter.py

@@ -2,6 +2,7 @@
 
 from guardrails.llm_providers import (
     ArbitraryCallable,
+    AsyncPromptCallableBase,
     PromptCallableBase,
 )
 
@@ -17,7 +18,15 @@ class BaseFormatter(ABC):
     @abstractmethod
     def wrap_callable(self, llm_callable: PromptCallableBase) -> ArbitraryCallable: ...
 
+    @abstractmethod
+    def wrap_async_callable(
+        self, llm_callable: PromptCallableBase
+    ) -> AsyncPromptCallableBase: ...
+
 
 class PassthroughFormatter(BaseFormatter):
-    def wrap_callable(self, llm_callable: PromptCallableBase):
+    def wrap_callable(self, llm_callable: PromptCallableBase):  # type: ignore
+        return llm_callable  # Noop
+
+    def wrap_async_callable(self, llm_callable: PromptCallableBase):  # type: ignore
         return llm_callable  # Noop

guardrails/formatters/json_formatter.py

@@ -149,3 +149,6 @@ def fn(
             raise ValueError(
                 "JsonFormatter can only be used with HuggingFace*Callable."
             )
+
+    def wrap_async_callable(self, llm_callable):
+        raise NotImplementedError()