Two model capabilities of AI big models: Function call and ReAct

In recent years , with the rapid development of large AI models, how to make these models better interact with the real world has become an important issue. Function Call and ReAct, as two important model capabilities, provide large models with more powerful tool calling and task execution capabilities. We will explore the background, principles, application scenarios and comparison of these two capabilities in depth to help you gain a deeper understanding of their value.

AI large models (such as GPT-4, PaLM, etc.) perform well in tasks such as natural language processing and text generation, but their capabilities are often limited to text generation and reasoning. In order to enable large models to interact with real-world tools, APIs, or external systems, researchers proposed two capabilities: Function Call and ReAct.

• Function Call: Allows large models to call external functions or APIs , thereby expanding their capabilities.

• ReAct: Combining reasoning and action, enabling the model to dynamically plan and execute multi-step operations in complex tasks .
  

  
  The emergence of these two capabilities marks the transformation of large models from "pure text generation" to "task execution".

2.1 Technical Principle

Function Call  is a capability that allows largemodels to call external functions or APIs . The core idea is to parse the output of the model into a specific function call request , which is then executed by the external system and the result is returned. The model can dynamically select the function to be called based on task requirements .

The specific process is as follows:

1. The model receives user input, performs intent recognition , and determines whether to call a function.

2. The model generates a function call request (including function name and parameters) based on the input content.

3. The external system executes the function and returns the result.

4. The model integrates the results and converts them into natural language for output .
   

   
   

   
   2.2 Code Examples
   

   
   The following is a simple Function Call example, assuming we have a weather query API :
   
   

# Define weather query function def get_weather(city: str): # Call weather API (assuming it has been implemented) return f"Weather in {city}: Sunny, 25°C" # Function call request generated by the model function_call = { "name": "get_weather", "parameters": {"city": "What's the weather like in Beijing"}} # Execute function call result = get_weather(**function_call["parameters"])print(result) # Output: Beijing: Sunny, 25°C

When the user enters "What's the weather like in Beijing", the model can automatically call the API, obtain weather information, and return the results.

2.3 Typical Application Scenarios

• Tool calls: such as querying the weather, translating text, calculating mathematical formulas, etc.

• Data acquisition: Get real-time data from a database or API.

• Automated tasks: such as sending emails, generating reports, etc.

3.1 Technical Principle

ReAct (Reasoning + Acting) is a task execution framework that combines reasoning and action . Unlike Function Call, ReAct not only calls tools, but also dynamically plans the task execution path through multi-step reasoning . The model will decide the next action based on the current state and environmental feedback .

ReAct's core process:

1. The model receives a task input.

2. The model performs inferences and generates the next action (such as calling a tool or answering a question).

3. Perform actions and get results.

4. Update the status based on the results and continue reasoning until the task is completed.
   

   
   3.2 Code Examples
   
   

The following is a simple ReAct example. Assume that the task is to query the weather in a city and recommend whether it is suitable for travel:

# Define tool function def get_weather(city: str): return f"Weather in {city}: Sunny, 25°C" # ReAct task execution def react_task(city: str): # Step 1: Get the weather weather = get_weather(city) print(weather) # Step 2: Infer whether it is suitable for travel if "Sunny" in weather: return f"It's a good day to visit {city}!" else: return f"Maybe reconsider visiting {city}." # Execute task result = react_task("Beijing")print(result) # Output: It's a good day to visit Beijing!

If you enter "北京", it will return: "Weather in Beijing: Sunny, 25°C".

In the second step, it will determine whether it is suitable for travel based on weather information:

• If the weather is sunny, it will tell you: "It's a good day to visit [city name]!"

• If the weather is not sunny, it will suggest you: "Maybe reconsider Then the overall code execution is:
  
  

• You tell the task: "Please help me determine whether Beijing is suitable for travel."

• The task will first obtain the weather in Beijing, and if it is sunny, it will tell you: "It's a good day to visit Beijing!" (Today is a good day to visit Beijing!).

3.3 Typical Application Scenarios

• Complex task planning: such as travel planning, project decomposition, etc.

• Dynamic decision making: Adjust action strategies based on environmental feedback.

• Multi-tool collaboration: Combine multiple tools to complete complex tasks.
  

  
  

Task: Based on the book title entered by the user, query the detailed information of the book and determine whether it is worth reading.

Use Function Call to implement:

step:

1. Call a "book query tool" to obtain detailed information about the book (such as rating, author, etc.) based on the title.

2. Based on the rating of the book, you can directly judge whether it is worth reading.

Running process:

• The user enters the title of the book: "The Little Prince".

• The system calls the "Book Query Tool" to obtain the rating of "The Little Prince" (for example, 4.5/5).

• The system directly judges: If the score is greater than 4, it is recommended for reading.

• Return result: "The Little Prince is rated 4.5, worth reading!"

Implemented using ReAct

step:

1. Infer user needs: The user wants to know whether "The Little Prince" is worth reading.

2. Call the "Book Query Tool" to obtain detailed information about "The Little Prince" (rating, author, introduction, etc.).

3. Whether the inference requires further analysis: For example, in addition to the rating, the author’s popularity should also be considered.

4. Call the "Author Information Tool" to check the popularity of the author of "The Little Prince" (Antoine de Saint-Exupéry).

5. Consider the comprehensive rating and author’s popularity to determine whether the book is worth reading.

6. Return result: "The Little Prince has a score of 4.5, and the author is the famous writer Saint-Exupéry. I highly recommend reading it!"

Obviously, ReAct performs better in complex tasks, while Function Call is more suitable for simple and direct calling scenarios.

Function Call and ReAct are two important capabilities of the AI ​​model, and they are applicable to different scenarios :

• Function Call  is suitable for simple, direct, and deterministic tasks .

• ReAct is suitable for complex, multi-step reasoning task planning . 
  
  

  
  

With the further development of AI technology , these two capabilities will bring a wider range of application scenarios to large models, from simple tool calls to complex task execution, and the capabilities of AI large models will continue to expand. In the future, we can expect more innovative applications that combine these two capabilities, bringing more possibilities to all walks of life.