microsoft/semantic-kernel · error · ServiceInvalidRequestError

Image content is not supported in a tool message.

Error message

Image content is not supported in a tool message.

What it means

Raised while formatting a tool-role message for Bedrock when an item is an ImageContent. The Bedrock tool-result converter accepts only TextContent and FunctionResultContent; image content in a tool result is explicitly rejected (the code comment notes image/document tool results are not yet supported by SK).

Source

Thrown at python/semantic_kernel/connectors/ai/bedrock/services/model_provider/utils.py:131

    return {
        "role": "assistant",
        "content": contents,
    }


def _format_tool_message(message: ChatMessageContent) -> dict[str, Any]:
    """Format a tool message to the expected object for the client.

    Args:
        message: The tool message.

    Returns:
        The formatted tool message.
    """
    contents: list[Any] = []
    for item in message.items:
        if isinstance(item, ImageContent):
            raise ServiceInvalidRequestError("Image content is not supported in a tool message.")

        if isinstance(item, TextContent):
            contents.append({"text": item.text})
        elif isinstance(item, FunctionResultContent):
            contents.append({
                "toolResult": {
                    "toolUseId": item.id,
                    # Image and document content are not yet supported in a tool message by SK
                    "content": [{"text": str(item)}],
                }
            })
        else:
            raise ServiceInvalidRequestError(f"Unsupported content type in a tool message: {type(item)}")

    return {
        "role": "user",
        "content": contents,
    }

View on GitHub (pinned to c028a0c7dc)

Solutions

1. Convert image function results to a text representation (e.g. a URL or description) before placing them in a tool message for Bedrock.
2. If the function must return binary image data, do not route it through Bedrock tool results; return metadata text instead.
3. Filter ImageContent out of tool messages before serialization when targeting Bedrock.

Example fix

// before
history.add_message(ChatMessageContent(role=AuthorRole.TOOL, items=[ImageContent(...)]))
// after
history.add_message(ChatMessageContent(role=AuthorRole.TOOL, items=[TextContent(text=str(image_url))]))

from semantic_kernel.contents.image_content import ImageContent

def tool_message_has_no_image(items) -> bool:
    return not any(isinstance(i, ImageContent) for i in items)

from semantic_kernel.contents.image_content import ImageContent

def is_bedrock_safe_tool_item(item: object) -> bool:
    return not isinstance(item, ImageContent)

from semantic_kernel.exceptions.service_exceptions import ServiceInvalidRequestError

try:
    await service.get_chat_message_contents(history, settings)
except ServiceInvalidRequestError as e:
    if "Image content is not supported in a tool message" in str(e):
        convert_image_tool_results_to_text(history)
    raise

When it happens

Trigger: A tool-role ChatMessageContent whose items include an ImageContent is sent through _format_tool_message to a Bedrock chat completion service.

Common situations: A plugin returns an image as its function result and it lands in a tool-role message; deserializing a stored tool message that carried image content; a vision-returning function being auto-invoked during function calling.

Related errors

• Image content is not supported in an assistant message.
• Unsupported content type in a tool message: {type(item)}
• Only text and image content are supported in a user message.
• Unsupported content type in an assistant message: {type(item
• No function result provided in the tool message.

AI-assisted analysis of microsoft/semantic-kernel@c028a0c7dc (2026-08-13). Data as JSON: /api/errors/f7376c0729468242. Report an issue: GitHub.