microsoft/autogen · error · ValueError

Unknown content type: {part}

Error message

Unknown content type: {part}

What it means

In _ollama_client.py's message conversion, each part of a list-content user message must be either a str or an autogen_core Image. Any other object (float, dict, a different image class, None) hits the else and raises ValueError with the offending part. The check is per-part, so a single bad element poisons the whole message.

Source

Thrown at python/packages/autogen-ext/src/autogen_ext/models/ollama/_ollama_client.py:187

                # name=message.source, # TODO: No name parameter in Ollama
            )
        ]
    else:
        ollama_messages: List[Message] = []
        for part in message.content:
            if isinstance(part, str):
                ollama_messages.append(Message(content=part, role="user"))
            elif isinstance(part, Image):
                # TODO: should images go into their own message? Should each image get its own message?
                if not ollama_messages:
                    ollama_messages.append(Message(role="user", images=[OllamaImage(value=part.to_base64())]))
                else:
                    if ollama_messages[-1].images is None:
                        ollama_messages[-1].images = [OllamaImage(value=part.to_base64())]
                    else:
                        ollama_messages[-1].images.append(OllamaImage(value=part.to_base64()))  # type: ignore
            else:
                raise ValueError(f"Unknown content type: {part}")
        return ollama_messages


def system_message_to_ollama(message: SystemMessage) -> Message:
    return Message(
        content=message.content,
        role="system",
    )


def _func_args_to_ollama_args(args: str) -> Dict[str, Any]:
    return json.loads(args)  # type: ignore


def func_call_to_ollama(message: FunctionCall) -> Message.ToolCall:
    return Message.ToolCall(
        function=Message.ToolCall.Function(
            name=message.name,

View on GitHub (pinned to 027ecf0a37)

Solutions

1. Wrap binary images with autogen_core Image: Image.from_pil(pil_img) or Image.from_file('x.png')
2. Filter/normalize parts before sending: keep only str and Image instances
3. Replace OpenAI-style part dicts with the corresponding str/Image values

Example fix

# before
content = ["describe", {"type": "image_url", "image_url": {"url": uri}}]
msg = UserMessage(content=content, source="user")

# after
from autogen_core.models import Image
content = ["describe", Image.from_file(image_path)]
msg = UserMessage(content=content, source="user")

from autogen_core.models import Image

def sanitize_parts(parts: Sequence[object]) -> list[str | Image]:
    out: list[str | Image] = []
    for p in parts:
        if isinstance(p, (str, Image)):
            out.append(p)
        else:
            raise TypeError(f"Unsupported content part: {type(p)!r}")
    return out

msg = UserMessage(content=sanitize_parts(raw_parts), source="user")

def is_valid_content_part(p: object) -> TypeGuard[str | Image]:
    return isinstance(p, (str, Image))

try:
    result = await client.create([msg])
except ValueError as e:
    if "Unknown content type" in str(e):
        msg = UserMessage(content=[p for p in msg.content if is_valid_content_part(p)], source="user")
        result = await client.create([msg])
    else:
        raise

When it happens

Trigger: UserMessage(content=['hi', 123]) or content=['text', {'type': 'image_url', ...}] (raw OpenAI-style part dicts); mixing in a PIL.Image or numpy array instead of autogen_core.models.Image; None entries from conditional part building.

Common situations: Translating OpenAI multi-part format directly instead of using Image.from_pil()/from_file; building content lists with appends that sometimes append falsy/None; a different library's Image type leaking into the message.

Related errors

• Model does not support vision and image was provided
• Multi-part messages such as those containing images are curr
• Model does not support vision and image was provided
• Invalid aggregate message {reason}
• The from field must be null or the agent name

AI-assisted analysis of microsoft/autogen@027ecf0a37 (2026-08-15). Data as JSON: /api/errors/c726d3bbe979b47d. Report an issue: GitHub.