semantic-kernel/python/notebooks/08-native-function-inline.ipynb

Running Native Functions

Two of the previous notebooks showed how to execute semantic functions inline and how to run prompts from a file.

In this notebook, we'll show how to use native functions from a file. We will also show how to call semantic functions from native functions.

This can be useful in a few scenarios:

• Writing logic around how to run a prompt that changes the prompt's outcome.
• Using external data sources to gather data to concatenate into your prompt.
• Validating user input data prior to sending it to the LLM prompt.

Native functions are defined using standard Python code. The structure is simple, but not well documented at this point.

The following examples are intended to help guide new users towards successful native & semantic function use with the SK Python framework.

Prepare a semantic kernel instance first, loading also the AI service settings defined in the Setup notebook:

In [ ]:

!python -m pip install semantic-kernel==0.9.2b1

In [ ]:

from services import Service

# Select a service to use for this notebook (available services: OpenAI, AzureOpenAI, HuggingFace)
selectedService = Service.OpenAI

In [ ]:

import semantic_kernel as sk
import semantic_kernel.connectors.ai.open_ai as sk_oai
from semantic_kernel.prompt_template.input_variable import InputVariable

kernel = sk.Kernel()

if selectedService == Service.AzureOpenAI:
    from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion

    deployment, api_key, endpoint = sk.azure_openai_settings_from_dot_env()
    service_id = "aoai_chat"  # used later in the notebook
    azure_chat_service = AzureChatCompletion(
        service_id=service_id, deployment_name="gpt-35-turbo", endpoint=endpoint, api_key=api_key
    )  # set the deployment name to the value of your chat model
    kernel.add_service(azure_chat_service)

# Configure OpenAI service
if selectedService == Service.OpenAI:
    from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

    api_key, org_id = sk.openai_settings_from_dot_env()
    service_id = "oai_chat"  # used later in the notebook
    oai_chat_service = OpenAIChatCompletion(
        service_id=service_id, ai_model_id="gpt-4-turbo-1106", api_key=api_key, org_id=org_id
    )
    kernel.add_service(oai_chat_service)

Let's create a native function that gives us a random number between 3 and a user input as the upper limit. We'll use this number to create 3-x paragraphs of text when passed to a semantic function.

First, let's create our native function.

In [ ]:

import random
from semantic_kernel.functions import kernel_function


class GenerateNumberPlugin:
    """
    Description: Generate a number between 3-x.
    """

    @kernel_function(
        description="Generate a random number between 3-x",
        name="GenerateNumberThreeOrHigher",
    )
    def generate_number_three_or_higher(self, input: str) -> str:
        """
        Generate a number between 3-<input>
        Example:
            "8" => rand(3,8)
        Args:
            input -- The upper limit for the random number generation
        Returns:
            int value
        """
        try:
            return str(random.randint(3, int(input)))
        except ValueError as e:
            print(f"Invalid input {input}")
            raise e

Next, let's create a semantic function that accepts a number as {{$input}} and generates that number of paragraphs about two Corgis on an adventure. $input is a default variable semantic functions can use.

In [ ]:

prompt = """
Write a short story about two Corgis on an adventure.
The story must be:
- G rated
- Have a positive message
- No sexism, racism or other bias/bigotry
- Be exactly {{$input}} paragraphs long. It must be this length.
"""

if selectedService == Service.OpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id="gpt-3.5-turbo-1106",
        max_tokens=2000,
        temperature=0.7,
    )
elif selectedService == Service.AzureOpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id=deployment,
        max_tokens=2000,
        temperature=0.7,
    )

prompt_template_config = sk.PromptTemplateConfig(
    template=prompt,
    name="story",
    template_format="semantic-kernel",
    input_variables=[
        InputVariable(name="input", description="The user input", is_required=True),
    ],
    execution_settings=execution_settings,
)

corgi_story = kernel.create_function_from_prompt(
    function_name="CorgiStory",
    plugin_name="CorgiPlugin",
    prompt_template_config=prompt_template_config,
)

generate_number_plugin = kernel.import_plugin_from_object(GenerateNumberPlugin(), "GenerateNumberPlugin")

In [ ]:

# Run the number generator
generate_number_three_or_higher = generate_number_plugin["GenerateNumberThreeOrHigher"]
number_result = await generate_number_three_or_higher(kernel, input=6)
print(number_result)

In [ ]:

story = await corgi_story.invoke(kernel, input=number_result.value)

Note: depending on which model you're using, it may not respond with the proper number of paragraphs.

In [ ]:

print(f"Generating a corgi story exactly {number_result.value} paragraphs long.")
print("=====================================================")
print(story)

Kernel Functions with Annotated Parameters

That works! But let's expand on our example to make it more generic.

For the native function, we'll introduce the lower limit variable. This means that a user will input two numbers and the number generator function will pick a number between the first and second input.

We'll make use of the Python's Annotated class to hold these variables.

In [ ]:

import sys
import semantic_kernel as sk
from semantic_kernel.connectors.ai.open_ai import (
    AzureChatCompletion,
    OpenAIChatCompletion,
)

if sys.version_info >= (3, 9):
    from typing import Annotated
else:
    from typing_extensions import Annotated

kernel = sk.Kernel()

if selectedService == Service.AzureOpenAI:
    from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion

    deployment, api_key, endpoint = sk.azure_openai_settings_from_dot_env()
    service_id = "aoai_chat"  # used later in the notebook
    azure_chat_service = AzureChatCompletion(
        service_id=service_id, deployment_name="gpt-35-turbo", endpoint=endpoint, api_key=api_key
    )  # set the deployment name to the value of your chat model
    kernel.add_service(azure_chat_service)

# Configure OpenAI service
if selectedService == Service.OpenAI:
    from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

    api_key, org_id = sk.openai_settings_from_dot_env()
    service_id = "oai_chat"  # used later in the notebook
    oai_chat_service = OpenAIChatCompletion(
        service_id=service_id, ai_model_id="gpt-4-turbo-1106", api_key=api_key, org_id=org_id
    )
    kernel.add_service(oai_chat_service)

Let's start with the native function. Notice that we're add the @kernel_function decorator that holds the name of the function as well as an optional description. The input parameters are configured as part of the function's signature, and we use the Annotated type to specify the required input arguments.

In [ ]:

import random
from semantic_kernel.functions import kernel_function


class GenerateNumberPlugin:
    """
    Description: Generate a number between a min and a max.
    """

    @kernel_function(
        name="GenerateNumber",
        description="Generate a random number between min and max",
    )
    def generate_number(
        self,
        min: Annotated[int, "the minimum number of paragraphs"],
        max: Annotated[int, "the maximum number of paragraphs"] = 10,
    ) -> Annotated[int, "the output is a number"]:
        """
        Generate a number between min-max
        Example:
            min="4" max="10" => rand(4,8)
        Args:
            min -- The lower limit for the random number generation
            max -- The upper limit for the random number generation
        Returns:
            int value
        """
        try:
            return str(random.randint(min, max))
        except ValueError as e:
            print(f"Invalid input {min} and {max}")
            raise e

In [ ]:

generate_number_plugin = kernel.import_plugin_from_object(GenerateNumberPlugin(), "GenerateNumberPlugin")
generate_number = generate_number_plugin["GenerateNumber"]

Now let's also allow the semantic function to take in additional arguments. In this case, we're going to allow the our CorgiStory function to be written in a specified language. We'll need to provide a paragraph_count and a language.

In [ ]:

prompt = """
Write a short story about two Corgis on an adventure.
The story must be:
- G rated
- Have a positive message
- No sexism, racism or other bias/bigotry
- Be exactly {{$paragraph_count}} paragraphs long
- Be written in this language: {{$language}}
"""

if selectedService == Service.OpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id="gpt-3.5-turbo-1106",
        max_tokens=2000,
        temperature=0.7,
    )
elif selectedService == Service.AzureOpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id=deployment,
        max_tokens=2000,
        temperature=0.7,
    )

prompt_template_config = sk.PromptTemplateConfig(
    template=prompt,
    name="summarize",
    template_format="semantic-kernel",
    input_variables=[
        InputVariable(name="paragraph_count", description="The number of paragraphs", is_required=True),
        InputVariable(name="language", description="The language of the story", is_required=True),
    ],
    execution_settings=execution_settings,
)

corgi_story = kernel.create_function_from_prompt(
    function_name="CorgiStory",
    plugin_name="CorgiPlugin",
    prompt_template_config=prompt_template_config,
)

Let's generate a paragraph count.

In [ ]:

result = await generate_number.invoke(kernel, min=1, max=5)
num_paragraphs = result.value
print(f"Generating a corgi story {num_paragraphs} paragraphs long.")

We can now invoke our corgi_story function using the kernel and the keyword arguments paragraph_count and language.

In [ ]:

# Pass the output to the semantic story function
desired_language = "Spanish"
story = await corgi_story.invoke(kernel, paragraph_count=num_paragraphs, language=desired_language)

In [ ]:

print(f"Generating a corgi story {num_paragraphs} paragraphs long in {desired_language}.")
print("=====================================================")
print(story)

Calling Native Functions within a Semantic Function

One neat thing about the Semantic Kernel is that you can also call native functions from within Prompt Functions!

We will make our CorgiStory semantic function call a native function GenerateNames which will return names for our Corgi characters.

We do this using the syntax {{plugin_name.function_name}}. You can read more about our prompte templating syntax here.

In [ ]:

import random
from semantic_kernel.functions import kernel_function


class GenerateNamesPlugin:
    """
    Description: Generate character names.
    """

    # The default function name will be the name of the function itself, however you can override this
    # by setting the name=<name override> in the @kernel_function decorator. In this case, we're using
    # the same name as the function name for simplicity.
    @kernel_function(description="Generate character names", name="generate_names")
    def generate_names(self) -> str:
        """
        Generate two names.
        Returns:
            str
        """
        names = {"Hoagie", "Hamilton", "Bacon", "Pizza", "Boots", "Shorts", "Tuna"}
        first_name = random.choice(list(names))
        names.remove(first_name)
        second_name = random.choice(list(names))
        return f"{first_name}, {second_name}"

In [ ]:

generate_names_plugin = kernel.import_plugin_from_object(GenerateNamesPlugin(), plugin_name="GenerateNames")
generate_names = generate_names_plugin["generate_names"]

In [ ]:

prompt = """
Write a short story about two Corgis on an adventure.
The story must be:
- G rated
- Have a positive message
- No sexism, racism or other bias/bigotry
- Be exactly {{$paragraph_count}} paragraphs long
- Be written in this language: {{$language}}
- The two names of the corgis are {{GenerateNames.generate_names}}
"""

In [ ]:

if selectedService == Service.OpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id="gpt-3.5-turbo-1106",
        max_tokens=2000,
        temperature=0.7,
    )
elif selectedService == Service.AzureOpenAI:
    execution_settings = sk_oai.OpenAIChatPromptExecutionSettings(
        service_id=service_id,
        ai_model_id=deployment,
        max_tokens=2000,
        temperature=0.7,
    )

prompt_template_config = sk.PromptTemplateConfig(
    template=prompt,
    name="corgi-new",
    template_format="semantic-kernel",
    input_variables=[
        InputVariable(name="paragraph_count", description="The number of paragraphs", is_required=True),
        InputVariable(name="language", description="The language of the story", is_required=True),
    ],
    execution_settings=execution_settings,
)

corgi_story = kernel.create_function_from_prompt(
    function_name="CorgiStoryUpdated",
    plugin_name="CorgiPluginUpdated",
    prompt_template_config=prompt_template_config,
)

In [ ]:

result = await generate_number.invoke(kernel, min=1, max=5)
num_paragraphs = result.value

In [ ]:

desired_language = "French"
story = await corgi_story.invoke(kernel, paragraph_count=num_paragraphs, language=desired_language)

In [ ]:

print(f"Generating a corgi story {num_paragraphs} paragraphs long in {desired_language}.")
print("=====================================================")
print(story)

Recap

A quick review of what we've learned here:

• We've learned how to create native and prompt functions and register them to the kernel
• We've seen how we can use Kernel Arguments to pass in more custom variables into our prompt
• We've seen how we can call native functions within a prompt.