How Can I Combine Function Calls with Descriptive Text in a Single API Response?

Below is an explanation of the current behavior and the recommended approach.

Current Behavior

When you instruct the model to make a function call (or when the model decides one is needed), the response is structured so that the function call is the primary output. In other words, the message’s content field is typically set to null (or is not used) and the function call information is returned instead. This means you won’t naturally get both a descriptive text and a function call in the same message.

Best Practice

Since you can’t directly mix a detailed text response with a designated function call in a single API response, the common approach is to split the conversation into two turns:

1. Function Call Message:
   The initial response from the model includes the function call (with its arguments) without additional text (content is null). This forces the function to be executed.

2. Follow-Up Response:
   Once your system executes the function and returns the result back to the conversation (adding the function’s output message to the conversation history), you then allow the model to generate a final reply that combines the initial narration with the function results. For example, the model can then produce text such as "I'm doing great thanks for asking, and the current time is 10:50 AM" (where the time comes from the function call).

This two-step conversational flow lets you maintain the clarity of the function call mechanism while also providing a narrative answer to the user.

Example Workflow

PYTHONCopied
# First API call: the assistant returns a function call (e.g. get_time)
messages = [
    {"role": "user", "content": "How are you doing today?"},
]

response = client.chat.completions.create(
    model="gpt-4o",
    messages=messages,
    tools=tools  # assuming get_time is configured as a tool
)

# The response typically looks like this:
# {
#   "message": {
#       "content": None,
#       "role": "assistant",
#       "function_call": {"name": "get_time", "arguments": "{}"}
#   }
# }

# Your system now executes the function get_time and gets, say, "10:50 AM"

# Second API call: include the function result in a new conversation message
messages.append({
    "role": "tool",
    "content": json.dumps("10:50 AM"),
    "tool_call_id": response.message.function_call.id  # reference the specific call if needed
})
messages.append({
    "role": "user",
    "content": "Please provide a final response that mentions your state and the fetched time."
})

final_response = client.chat.completions.create(
    model="gpt-4o",
    messages=messages,
    tools=tools
)

# The final_response might have:
# content: "I'm doing great thanks for asking, and the current time is 10:50 AM"

Summary

• Direct Combination: It isn’t possible to get both text content and a function call in the same API response.
• Recommended Pattern: Use a two-turn process where the function call is handled first and its result provided to the model in a subsequent conversation turn. This allows the model to craft a fully contextualized answer combining both the narrative and data from the function call.

This workflow helps you achieve the desired combined outcome while adhering to the design of the API’s function calling mechanism.