microsoft/semantic-kernel · error · ServiceInvalidResponseError

Invalid response type from Ollama chat completion. Expected

Error message

Invalid response type from Ollama chat completion. Expected AsyncIterator but got {type(response_object)}.

What it means

Raised inside _inner_get_streaming_text_contents after a streaming call to self.client.generate(stream=True). When stream=True the ollama client should return an AsyncIterator of generate chunks; if it returns a single object instead, the connector cannot iterate and raises ServiceInvalidResponseError. Like error 1121, the message says 'chat completion' but this is the streaming text-completion path.

Source

Thrown at python/semantic_kernel/connectors/ai/ollama/services/ollama_text_completion.py:148

    @trace_streaming_text_completion(OllamaBase.MODEL_PROVIDER_NAME)
    async def _inner_get_streaming_text_contents(
        self,
        prompt: str,
        settings: "PromptExecutionSettings",
    ) -> AsyncGenerator[list[StreamingTextContent], Any]:
        if not isinstance(settings, OllamaTextPromptExecutionSettings):
            settings = self.get_prompt_execution_settings_from_settings(settings)
        assert isinstance(settings, OllamaTextPromptExecutionSettings)  # nosec

        response_object = await self.client.generate(
            model=self.ai_model_id,
            prompt=prompt,
            stream=True,
            **settings.prepare_settings_dict(),
        )

        if not isinstance(response_object, AsyncIterator):
            raise ServiceInvalidResponseError(
                "Invalid response type from Ollama chat completion. "
                f"Expected AsyncIterator but got {type(response_object)}."
            )

        async for part in response_object:
            yield [
                StreamingTextContent(
                    choice_index=0,
                    inner_content=part,
                    ai_model_id=self.ai_model_id,
                    text=part.response if isinstance(part, GenerateResponse) else part.get("response"),
                )
            ]

    # endregion

View on GitHub (pinned to c028a0c7dc)

Solutions

Ensure you are using a compatible ollama Python package version that returns an AsyncIterator for stream=True
If passing a custom client, make sure it returns an async generator from generate(stream=True)
In tests, mock self.client.generate to return an async iterator, not a single object
Fall back to non-streaming get_text_contents if streaming is unreliable with your setup

Defensive patterns

Strategy: try-catch

Type guard

from collections.abc import AsyncIterator

def is_streaming_response(obj) -> bool:
    return isinstance(obj, AsyncIterator)

Try / catch

from semantic_kernel.exceptions.service_exceptions import ServiceInvalidResponseError

try:
    async for chunk in ollama.get_streaming_text_contents(prompt='Hello', settings=settings):
        process(chunk)
except ServiceInvalidResponseError as e:
    logger.error('Ollama streaming returned unexpected type: %s', e)
    # fall back to non-streaming
    result = await ollama.get_text_contents(prompt='Hello', settings=settings)

Prevention

Ensure your ollama client version supports async streaming generators
When mocking the client in tests, return an async generator from generate(stream=True)
Provide a non-streaming fallback path for environments where streaming is unreliable

When it happens

Trigger: Calling get_streaming_text_contents when the ollama client ignores stream=True and returns a single GenerateResponse — happens with certain ollama client versions, with mock/stub clients in tests, or when a custom client passed to the constructor does not support async iteration.

Common situations: Testing with a fake AsyncClient that returns a dict instead of an async generator; ollama library version mismatch where streaming semantics changed; a custom client object that does not implement __aiter__.

Related errors

AI-assisted analysis of microsoft/semantic-kernel@c028a0c7dc (2026-08-13). Data as JSON: /api/errors/0a87ed7f88ccdd38. Report an issue: GitHub.