Python: Bug: default_services in config.json for plugin does not work

[xyang2013]

Model "ollama-llama3.1:8b" cannot be selected according to config file. The following code was executed in the latest version of VS Code.

sample code

import semantic_kernel as sk

kernel = sk.Kernel()

from semantic_kernel.connectors.ai.ollama import OllamaChatCompletion

kernel.add_service(
    OllamaChatCompletion(
        service_id="ollama-qwen3:30b",
        host="http://localhost:11434",
        ai_model_id="qwen3:30b",
    )
)

kernel.add_service(
    OllamaChatCompletion(
        service_id="ollama-llama3.1:8b",
        host="http://localhost:11434",
        ai_model_id="llama3.1:8b",
    )
)

jokes_plugin = kernel.add_plugin(None, parent_directory="plugins", plugin_name="jokes")

theme = "Dishes"

from semantic_kernel.functions.kernel_arguments import KernelArguments

knock_joke = await kernel.invoke(jokes_plugin["knock_knock_joke"], KernelArguments(input=theme))
print(knock_joke)

config.json

{
    "schema": 1,
    "type": "completion",
    "description": "Generates a knock-knock joke based on user input",
    "default_services": [
        "ollama-llama3.1:8b"
    ],
    "execution_settings": {
        "default": {
            "temperature": 0.8,
            "number_of_responses": 1,
            "top_p": 1,
            "max_tokens": 4000,
            "presence_penalty": 0.0,
            "frequency_penalty": 0.0
        }
    },
    "input_variables": [
        {
            "name": "input",
            "description": "The topic that the joke should be written about",
            "required": true
        }
    ]
}

[moonbox3]

Hi @xyang2013, in SK Python we don't have anything related to default_services as part of the config.json file. The config.json is used to manage the execution settings for a prompt function. One can have a dict of service_id to execution_settings so they can easily load and then switch between the settings for different service ids. You'll need to do something like this to create the AI service:

OllamaChatCompletion(ai_model_id="ollama-llama3.1:8b")

If you specify the service_id as part of the OllamaChatCompletion(...) constructor, then you will also configure that same service_id in your config.json. Otherwise, it gets set to default if it isn't explicitly configured.

[xyang2013]

Hi @moonbox3. In the above code, I already created two services with different service_id s:

kernel.add_service(
    OllamaChatCompletion(
        service_id="ollama-qwen3:30b",
        host="http://localhost:11434",
        ai_model_id="qwen3:30b",
    )
)

kernel.add_service(
    OllamaChatCompletion(
        service_id="ollama-llama3.1:8b",
        host="http://localhost:11434",
        ai_model_id="llama3.1:8b",
    )
)

How do you specify the service_id in the execution_settings?

"execution_settings": {
        "default": {
            "temperature": 0.8,
            "number_of_responses": 1,
            "top_p": 1,
            "max_tokens": 4000,
            "presence_penalty": 0.0,
            "frequency_penalty": 0.0
        }
    },

By the way, default_services use case was from a book example. But in the example, the author used ChatGPT. Here, I tried to use local LLMs.

Thank you.

[moonbox3]

@xyang2013 thanks for your response. Your execution setting's dict only has the default key, but then you're creating the AI Services with either ollama-qwen3:30b or ollama-llama3.1:8b. You will simply do this:

"execution_settings": {
        "ollama-qwen3:30b": {
            "temperature": 0.8,
            "number_of_responses": 1,
            "top_p": 1,
            "max_tokens": 4000,
            "presence_penalty": 0.0,
            "frequency_penalty": 0.0
        },
       "ollama-llama3.1:8b": {
            "temperature": 0.2,
            "number_of_responses": 1,
            "top_p": 1,
            "max_tokens": 2000,
            "presence_penalty": 0.0,
            "frequency_penalty": 0.0
        }
    },

Which book did this come from? Is it a Semantic Kernel .NET-related book? SK .NET may handle this slightly differently than SK Python FYI.

[xyang2013]

@moonbox3 Thank you for your suggestions. I can now specify which local model to target in the JSON file.

The book I was referring to is "Building AI Applications with Microsoft Semantic Kernel: Easily integrate generative AI capabilities and copilot experiences into your applications."

By the way, is there a debug mode similar to what Langchain offers? I find it hard to see what's happening under the hood during complex invokes, like when using the planner.