Inconsistent Attachment Handling Between MCP and semantic-kernel MCP Server

[nblog]

The issue described here is similar to #12850.

---

Problem Description

I am trying to migrate functionality from an MCP server implementation to semantic-kernel, specifically regarding returning attachments as part of a function response. My implementation works as expected with MCP, but fails to produce the desired result when using semantic-kernel.

MCP Example (Works as Intended)

@mcp.tool()
def analyze_logs(log_path: str):
    return [
        "分析附件中的日志文件，并解释错误原因有哪些。",
        BlobResourceContents(
            uri=log_path,
            blob=base64.b64encode(open(log_path, mode="rb").read()).decode(),
            mimeType="text/plain",
        )
    ]

semantic-kernel Example (Does Not Work)

@kernel_function()
def analyze_logs(log_path: str):
    return [
        TextContent(text="分析附件中的日志文件，并解释错误原因有哪些。"),
        BinaryContent.from_file(log_path)
    ]

---

Request for Guidance

• Is this problem caused by a difference in how attachments (such as binary or file resources) are transmitted in semantic-kernel compared to MCP?
• If so, could you clarify the correct approach for returning attachments or binary resources from a kernel function in semantic-kernel’s MCP server?
• If the issue is with my code, please point out the necessary corrections.

Any guidance or best practices for passing attachments through semantic-kernel MCP server functions would be greatly appreciated.

Thank you!

[nblog]

@eavanvalkenburg could you take a look at it?

[eavanvalkenburg]

@nblog python version 1.35.2 was just released and has a fix for this, I made it for the related discussion you flagged, but it improves the parsing of types across the board so should also do something here. Please have a test and let me know!

[nblog]

@eavanvalkenburg

I have upgraded to the latest version of semantic-kernel (1.35.2) for validation, but the issue described above still persists.

---

Details

1. Behavior in MCP (BlobResourceContents):

   The response for BlobResourceContents data is as follows (I also find it odd that my binary data is wrapped as text):

   {
     "content": [
       {
         "type": "text",
         "text": "分析附件中的PDF文件。"
       },
       {
         "type": "text",
         "text": "{\n  \"uri\": \"c:\\\\Public\\\\PDF\\\\asm2vec.pdf\",\n  \"mimeType\": \"application/pdf\",\n  \"_meta\": null,\n  \"blob\": \"PyFLtsR...\""
       }
     ],
     "isError": false
   }

2. Behavior in Semantic Kernel (BinaryContent):

   In contrast, the response for BinaryContent data in semantic-kernel is as follows (note that the type has changed to resource, which is likely the reason why it does not work properly in VS Code):

   {
     "content": [
       {
         "type": "text",
         "text": "分析附件中的PDF文件"
       },
       {
         "type": "resource",
         "resource": {
           "uri": "c:\\Public\\PDF\\asm2vec.pdf",
           "mimeType": "text/plain",
           "blob": "JVBERi0.."
         }
       }
     ],
     "isError": false
   }

---

Additional Notes

• The inconsistency in response format—specifically, the use of type: "resource" and the altered mimeType—may be causing the integration issue with VS Code.
• It is worth noting that the binary data is handled differently in each case, which could lead to compatibility problems.

---

MCP (Working):

Semantic-Kernel:

[nblog]

I encountered a problem when using the link address directly, specifically with the following usage pattern:

As shown in the official example python/samples/concepts/chat_completion/simple_chatbot_with_image.py , the following code demonstrates that ImageContent(uri=IMAGE_URI) should support image URLs:

image_content_remote = ImageContent(uri=IMAGE_URI)

However, in my own implementation, I attempted the following:

@kernel_function()
def my_tool():
  contents = [TextContent(text="summarize these points")]
  # image
  contents.append(ImageContent(uri=img_info['url']))
  # log file
  contents.append(BinaryContent(uri=file_info['url'], mime_type=file_info['mime_type']))
  return ChatMessageContent(role="tool", items=contents)

When I inspect the response in MCP, I find that the data field contains no content.

---

I am not sure whether this issue is caused by the internal implementation of semantic-kernel or by the MCP wrapper, but I hope this can be addressed and resolved.