Guide to AI Prompt and Tool Use (Function Calling) via API

TLDR;

To achieve the best results with AI models, you need to understand how to craft effective prompts and utilize tool use (also known as function calling) efficiently. These are the best practices to follow:

Define AI Role and Purpose
Contextual and specific
Give AI Examples
Set Parameters for the AI response
Use Structured Outputs

Crafting Effective AI Prompts

Writing effective prompt and utilizing tool use (also known as function calling) efficiently are essential skills when working with AI models. Whether you’re working with AI for content generation or creating custom AI tools, crafting well-structured prompts and understanding tool use can significantly improve the accuracy and relevance of AI responses. This guide will help you master these techniques.

Here are some best practices for crafting effective AI prompts.

Define AI Role and Purpose

Set AI role helps it generate more accurate and relevant responses. Clearly define what kind of assistant you need.

- You are a helpful customer service assistant.
+ You are an AI customer support assistant for an e-commerce store. You will be guiding the customer needs.

Contextual and specific

Provide the AI with the necessary context and specific information to perform the task.

You are an AI customer support assistant for an e-commerce store. You will be guiding the customer needs.

- You need to assist customers with things like faq, return, exchange, feedback, complaint, tracking, etc.
+ You will handle various customer service scenarios, including FAQs, returns, exchanges, feedback, complaints, and order tracking. For each scenario, provide a helpful and professional response based on the customer's inquiry.

Give AI Examples

Provide the AI with explanations or examples for input or output.

Example:

You will handle various customer service scenarios, including FAQs, returns, exchanges, feedback, complaints, and order tracking. For each scenario, provide a helpful and professional response based on the customer's inquiry.

+ Return:
+ Any questions related to the return process and policy
+
+ Exchange:
+ Any questions related to the exchange process and policy
+
+ Feedback:
+ Any questions related to the feedback process and policy
+ Review or feedback on the product or service
+
+ Complaint:
+ Bad experience with the product or service
+
+ Order Tracking:
+ Any questions related to the order tracking
+
+ FAQ:
+ Any 5W1H customer questions about the services that does not match the return, exchange, feedback, complaint, order tracking.

Set Parameters for the AI response

Setting constraints ensures AI provides responses that fit within desired parameters such as word limits, styles, or any other specific requirements.

Example:
```
Use maximum 100 words
Mantain formal and professional tone
Response in English
```

Use Structured Outputs

Use structured outputs to ensure the AI response is in the desired format to enhance clarity and usability. There are two ways to use structured outputs:

Inline Prompt
Structured Outputs

Using Inline Prompt:

Response in JSON format with the following properties:
- confidence: number between 0 and 1, your confidence score for the response.
- category: string, the category of the customer request.
- response: string, the response to the customer request.

Using Structured Outputs:

{
  "name": "customer_support_response",
  "type": "object",
  "properties": {
    "confidence": { 
     "type": "number",
     "description": "Your confidence score for the response between 0 and 1."
    },
    "category": { 
     "type": "string",
     "description": "The category of the customer request."
    },
    "response": { 
     "type": "string",
     "description": "The response to the customer request."
    }
  }
}

Crafting Effective Tool Use (Function Calling)

Tool use (also known as function calling) allows AI models to interact with external tools. This capability enables AI to retrieve real-time data, perform calculations, and execute predefined functions.

How Function Calling Works

AI Receives a Query

When a user makes a request, the AI first analyzes the query to determine if an external function is required to fetch accurate information.

Example:
```
I need to know my order status with order id 41KG76MN
```
AI Calls a Function

If necessary, the AI formulates a structured request to an external system, API, or database with relevant parameters.

Example:
```
{
 "name": "get_order_status",
 "parameters": {
   "order_id": "41KG76MN"
 }
}
```

Function Executes and Returns Data

The requested system processes the query and returns the needed information or action result.

Example:

Execute the function get_order_status with the parameters order_id: 41KG76MN

Function Response:

{
 "order_id": "41KG76MN",
 "delivery_status": "Shipped",
 "shipping_status": "With Courier",
 "shipping_date": "2025-02-10",
 "estimated_delivery_date": "2025-02-15"
}

AI Generates a Final Response

The AI integrates the functions output into a user-friendly response, ensuring clarity and completeness.

Example:
```
Your order #41KG76MN is shipped with courier and will be delivered on 2025-02-15.
```

Example of Prompt with Function Calling

Prompt:

You are an AI customer support assistant for an e-commerce store. Your task is to onboard the customer to the e-commerce store.

You need to collect the following information from the customer:
- Name
- Email
- Phone number

Rules:
1. Response in English.
2. Mantain formal and professional tone.
3. Always prioritize collecting the required information before proceeding with the inquiry. If the customer message does not contain the necessary details, ask for them politely before moving forward.

Function:

{
  "name": "onboard_customer",
  "description": "Customer onboarded to the e-commerce store",
  "parameters": {
    "type": "object",
    "properties": {
      "name": {
        "type": "string",
        "description": "Customer name"
      },
      "email": {
        "type": "string",
        "description": "Customer email"
      },
      "phone": {
        "type": "string",
        "description": "Customer phone number"
      },
    },
    "required": ["name", "email", "phone"],
    "additionalProperties": false
  }
}

If the customer provided all the informations, The AI model will return the following response:

{
  "name": "onboard_customer",
  "parameters": {
    "name": "John Doe",
    "email": "john.doe@example.com",
    "phone": "(123) 456-7890"
  }
}

You can modify the description so it will be more specific and clear for the AI model to understand the function.

{
  ...
  "phone": {
    "type": "string",
-   "description": "Customer phone number"
+   "description": "Customer phone number, convert to a plain numeric format"
  }
  ...
}

So the output will be:

{
  "name": "onboard_customer",
  "parameters": {
    "name": "John Doe",
    "email": "john.doe@example.com",
    "phone": "1234567890"
  }
}

Best Practices for Function Calling

Use descriptive names for functions and parameters.
For parameters with a fixed set of options, use enums to restrict input.
Sanitize and validate user-provided input before passing it to functions to prevent errors or security vulnerabilities.
Implement error handling for scenarios where Functions or API requests fail or return unexpected results.