← 课程目录 · 东灵培智
GitHub 原文 ↗ 中文
III · 构建 AI 应用

Integrating with function calling

微软《Generative AI for Beginners》 · 第 11 课 · 英文原版 · 本地镜像

Integrating with function calling

You've learned a fair bit so far in the previous lessons. However, we can improve further. Some things we can address are how we can get a more consistent response format to make it easier to work with the response downstream. Also, we might want to add data from other sources to further enrich our application.

The above-mentioned problems are what this chapter is looking to address.

Introduction

This lesson will cover:

Learning Goals

By the end of this lesson, you will be able to:

Scenario: Improving our chatbot with functions

For this lesson, we want to build a feature for our education startup that allows users to use a chatbot to find technical courses. We will recommend courses that fit their skill level, current role and technology of interest.

To complete this scenario, we will use a combination of:

To get started, let's look at why we would want to use function calling in the first place:

Why Function Calling

Before function calling, responses from an LLM were unstructured and inconsistent. Developers were required to write complex validation code to make sure they were able to handle each variation of a response. Users could not get answers like "What is the current weather in Stockholm?". This is because models were limited to the time the data was trained on.

Function Calling is a feature of the Azure OpenAI Service to overcome the following limitations:

Illustrating the problem through a scenario

We recommend you to use the included notebook if you want to run the below scenario. You can also just read along as we're trying to illustrate a problem where functions can help to address the problem.

Let's look at the example that illustrates the response format problem:

Let's say we want to create a database of student data so we can suggest the right course to them. Below we have two descriptions of students that are very similar in the data they contain.

  1. Create a connection to our Azure OpenAI resource:

```python import os import json from openai import OpenAI from dotenv import load_dotenv load_dotenv()

# The Responses API is served from the Azure OpenAI (Microsoft Foundry) v1 # endpoint, so we point the OpenAI client at /openai/v1/. endpoint = os.environ['AZURE_OPENAI_ENDPOINT'] client = OpenAI( api_key=os.environ['AZURE_OPENAI_API_KEY'], base_url=f"{endpoint.rstrip('/')}/openai/v1/", )

deployment=os.environ['AZURE_OPENAI_DEPLOYMENT'] ```

Below is some Python code for configuring our connection to Azure OpenAI. Because we use the v1 endpoint, we only need to set the api_key and base_url (no api_version is required).

  1. Creating two student descriptions using variables student_1_description and student_2_description.

```python student_1_description="Emily Johnson is a sophomore majoring in computer science at Duke University. She has a 3.7 GPA. Emily is an active member of the university's Chess Club and Debate Team. She hopes to pursue a career in software engineering after graduating."

student_2_description = "Michael Lee is a sophomore majoring in computer science at Stanford University. He has a 3.8 GPA. Michael is known for his programming skills and is an active member of the university's Robotics Club. He hopes to pursue a career in artificial intelligence after finishing his studies." ```

We want to send the above student descriptions to an LLM to parse the data. This data can later be used in our application and be sent to an API or stored in a database.

  1. Let's create two identical prompts in which we instruct the LLM on what information we are interested in:

```python prompt1 = f''' Please extract the following information from the given text and return it as a JSON object:

name major school grades club

This is the body of text to extract the information from: {student_1_description} '''

prompt2 = f''' Please extract the following information from the given text and return it as a JSON object:

name major school grades club

This is the body of text to extract the information from: {student_2_description} ''' ```

The above prompts instruct the LLM to extract information and return the response in JSON format.

  1. After setting up the prompts and the connection to Azure OpenAI, we will now send the prompts to the LLM by using client.responses.create. We store the prompt in the input variable and assign the role to user. This is to mimic a message from a user being written to a chatbot.

```python # response from prompt one openai_response1 = client.responses.create( model=deployment, input = [{'role': 'user', 'content': prompt1}], store=False, ) openai_response1.output_text

# response from prompt two openai_response2 = client.responses.create( model=deployment, input = [{'role': 'user', 'content': prompt2}], store=False, ) openai_response2.output_text ```

Now we can send both requests to the LLM and examine the response we receive by finding it like so openai_response1.output_text.

  1. Lastly, we can convert the response to JSON format by calling json.loads:

python # Loading the response as a JSON object json_response1 = json.loads(openai_response1.output_text) json_response1

Response 1:

json { "name": "Emily Johnson", "major": "computer science", "school": "Duke University", "grades": "3.7", "club": "Chess Club" }

Response 2:

json { "name": "Michael Lee", "major": "computer science", "school": "Stanford University", "grades": "3.8 GPA", "club": "Robotics Club" }

Even though the prompts are the same and the descriptions are similar, we see values of the Grades property formatted differently, as we can sometimes get the format 3.7 or 3.7 GPA for example.

This result is because the LLM takes unstructured data in the form of the written prompt and returns also unstructured data. We need to have a structured format so that we know what to expect when storing or using this data

So how do we solve the formatting problem then? By using functional calling, we can make sure that we receive structured data back. When using function calling, the LLM does not actually call or run any functions. Instead, we create a structure for the LLM to follow for its responses. We then use those structured responses to know what function to run in our applications.

function flow

We can then take what is returned from the function and send this back to the LLM. The LLM will then respond using natural language to answer the user's query.

Use Cases for using function calls

There are many different use cases where function calls can improve your app like:

Creating Your First Function Call

The process of creating a function call includes 3 main steps:

  1. Calling the Responses API with a list of your functions (tools) and a user message.
  2. Reading the model's response to perform an action i.e. execute a function or API Call.
  3. Making another call to the Responses API with the response from your function to use that information to create a response to the user.

LLM Flow

Step 1 - creating messages

The first step is to create a user message. This can be dynamically assigned by taking the value of a text input or you can assign a value here. If this is your first time working with the Responses API, we need to define the role and the content of the message.

The role can be either system (creating rules), assistant (the model) or user (the end-user). For function calling, we will assign this as user and an example question.

messages= [ {"role": "user", "content": "Find me a good course for a beginner student to learn Azure."} ]

By assigning different roles, it's made clear to the LLM if it's the system saying something or the user, which helps to build a conversation history that the LLM can build upon.

Step 2 - creating functions

Next, we will define a function and the parameters of that function. We will use just one function here called search_courses but you can create multiple functions.

Important : Functions are included in the system message to the LLM and will be included in the amount of available tokens you have available.

Below, we create the functions as an array of items. Each item is a tool in the flat Responses API format, with properties type, name, description and parameters:

functions = [
   {
      "type":"function",
      "name":"search_courses",
      "description":"Retrieves courses from the search index based on the parameters provided",
      "parameters":{
         "type":"object",
         "properties":{
            "role":{
               "type":"string",
               "description":"The role of the learner (i.e. developer, data scientist, student, etc.)"
            },
            "product":{
               "type":"string",
               "description":"The product that the lesson is covering (i.e. Azure, Power BI, etc.)"
            },
            "level":{
               "type":"string",
               "description":"The level of experience the learner has prior to taking the course (i.e. beginner, intermediate, advanced)"
            }
         },
         "required":[
            "role"
         ]
      }
   }
]

Let's describe each function instance more in detail below:

There's also an optional property required - required property for the function call to be completed.

Step 3 - Making the function call

After defining a function, we now need to include it in the call to the Responses API. We do this by adding tools to the request. In this case tools=functions.

There is also an option to set tool_choice to auto. This means we will let the LLM decide which function should be called based on the user message rather than assigning it ourselves.

Here's some code below where we call client.responses.create, note how we set tools=functions and tool_choice="auto" and thereby giving the LLM the choice when to call the functions we provide it:

response = client.responses.create(model=deployment,
                                        input=messages,
                                        tools=functions,
                                        tool_choice="auto",
                                        store=False)

print(response.output)

The response coming back now includes a function_call item in response.output that looks like so:

{
  "type": "function_call",
  "name": "search_courses",
  "call_id": "call_abc123",
  "arguments": "{\n  \"role\": \"student\",\n  \"product\": \"Azure\",\n  \"level\": \"beginner\"\n}"
}

Here we can see how the function search_courses was called and with what arguments, as listed in the arguments property in the JSON response.

The conclusion the LLM was able to find the data to fit the arguments of the function as it was extracting it from the value provided to the input parameter in the Responses API call. Below is a reminder of the messages value:

messages= [ {"role": "user", "content": "Find me a good course for a beginner student to learn Azure."} ]

As you can see, student, Azure and beginner was extracted from messages and set as input to the function. Using functions this way is a great way to extract information from a prompt but also to provide structure to the LLM and have reusable functionality.

Next, we need to see how we can use this in our app.

Integrating Function Calls into an Application

After we have tested the formatted response from the LLM, we can now integrate this into an application.

Managing the flow

To integrate this into our application, let's take the following steps:

  1. First, let's make the call to the OpenAI services and extract the function call items from the response output.

python response_items = response.output tool_calls = [item for item in response_items if item.type == "function_call"]

  1. Now we will define the function that will call the Microsoft Learn API to get a list of courses:

```python import requests

def search_courses(role, product, level): url = "https://learn.microsoft.com/api/catalog/" params = { "role": role, "product": product, "level": level } response = requests.get(url, params=params) modules = response.json()["modules"] results = [] for module in modules[:5]: title = module["title"] url = module["url"] results.append({"title": title, "url": url}) return str(results) ```

Note how we now create an actual Python function that maps to the function names introduced in the functions variable. We're also making real external API calls to fetch the data we need. In this case, we go against the Microsoft Learn API to search for training modules.

Ok, so we created functions variables and a corresponding Python function, how do we tell the LLM how to map these two together so our Python function is called?

  1. To see if we need to call a Python function, we need to look into the LLM response and see if a function_call item is part of it and call the pointed-out function. Here's how you can make the mentioned check below:

```python # Check if the model wants to call a function if tool_calls: for tool_call in tool_calls: print("Recommended Function call:") print(tool_call.name) print()

 # Call the function.
 function_name = tool_call.name

 available_functions = {
         "search_courses": search_courses,
 }
 function_to_call = available_functions[function_name]

 function_args = json.loads(tool_call.arguments)
 function_response = function_to_call(**function_args)

 print("Output of function call:")
 print(function_response)
 print(type(function_response))

 # Add the function call and its result back to the conversation.
 # The model's function_call item must be appended before its output.
 messages.append(tool_call)  # the assistant's function_call item
 messages.append( # the function result
     {
         "type": "function_call_output",
         "call_id": tool_call.call_id,
         "output": function_response,
     }
 )

```

These three lines, ensure we extract the function name, the arguments and make the call:

```python function_to_call = available_functions[function_name]

function_args = json.loads(tool_call.arguments) function_response = function_to_call(**function_args) ```

Below is the output from running our code:

Output

```Recommended Function call: { "name": "search_courses", "arguments": "{\n \"role\": \"student\",\n \"product\": \"Azure\",\n \"level\": \"beginner\"\n}" }

Output of function call: [{'title': 'Describe concepts of cryptography', 'url': 'https://learn.microsoft.com/training/modules/describe-concepts-of-cryptography/? WT.mc_id=api_CatalogApi'}, {'title': 'Introduction to audio classification with TensorFlow', 'url': 'https://learn.microsoft.com/en- us/training/modules/intro-audio-classification-tensorflow/?WT.mc_id=api_CatalogApi'}, {'title': 'Design a Performant Data Model in Azure SQL Database with Azure Data Studio', 'url': 'https://learn.microsoft.com/training/modules/design-a-data-model-with-ads/? WT.mc_id=api_CatalogApi'}, {'title': 'Getting started with the Microsoft Cloud Adoption Framework for Azure', 'url': 'https://learn.microsoft.com/training/modules/cloud-adoption-framework-getting-started/?WT.mc_id=api_CatalogApi'}, {'title': 'Set up the Rust development environment', 'url': 'https://learn.microsoft.com/training/modules/rust-set-up-environment/?WT.mc_id=api_CatalogApi'}] ```

  1. Now we will send the updated message, messages to the LLM so we can receive a natural language response instead of an API JSON formatted response.

```python print("Messages in next request:") print(messages) print()

second_response = client.responses.create( input=messages, model=deployment, tool_choice="auto", tools=functions, temperature=0, store=False, ) # get a new response from the model where it can see the function response

print(second_response.output_text) ```

Output

```text I found some good courses for beginner students to learn Azure:

  1. Describe concepts of cryptography
  2. Introduction to audio classification with TensorFlow
  3. Design a Performant Data Model in Azure SQL Database with Azure Data Studio
  4. Getting started with the Microsoft Cloud Adoption Framework for Azure
  5. Set up the Rust development environment

You can click on the links to access the courses. ```

Assignment

To continue your learning of Azure OpenAI Function Calling you can build:

Hint: Follow the Learn API reference documentation page to see how and where this data is available.

Great Work! Continue the Journey

After completing this lesson, check out our Generative AI Learning collection to continue leveling up your Generative AI knowledge!

Head over to Lesson 12, where we will look at how to design UX for AI applications!