tool_choice and control

Control the model’s behavior

By default, the model decides on its own whether to call a tool or respond directly. This is auto mode. But in some cases, you need finer control. The tool_choice parameter gives you that power.

The four tool_choice modes

auto mode (default)

response = client.chat.completions.create(
    model="grok-4.5",
    messages=messages,
    tools=tools,
    tool_choice="auto"
)

The model evaluates the request and decides:

• If the request requires external data → it calls a tool
• If it can respond directly → it responds without a tool

This is the most common and natural mode. Use it when you trust the model’s judgment.

required mode

response = client.chat.completions.create(
    model="grok-4.5",
    messages=messages,
    tools=tools,
    tool_choice="required"
)

The model must call at least one tool. It cannot respond directly, even if it already knows the answer. Use cases:

• Structured extraction: You want to force the model to extract data into a specific format via a dedicated tool
• Mandatory validation: Every request must pass through a verification tool
• Systematic logging: Everything must be traced via a logging tool

none mode

response = client.chat.completions.create(
    model="grok-4.5",
    messages=messages,
    tools=tools,
    tool_choice="none"
)

Completely disables function calling for this request. The model sees the tool definitions (they remain in context) but cannot call them. Use cases:

• Clarification phase: You want the model to ask questions before acting
• Preview mode: Show what the model would do without executing
• Fallback: If tools are temporarily unavailable

Specific mode (force a specific tool)

response = client.chat.completions.create(
    model="grok-4.5",
    messages=messages,
    tools=tools,
    tool_choice={
        "type": "function",
        "function": {"name": "get_weather"}
    }
)

The model is forced to call exactly this tool. It cannot choose another one, nor respond directly. This is the most restrictive mode. Use cases:

• Sequential workflow: At this step, this specific tool must be called
• Tests: You want to verify that the model generates the correct arguments for a given tool
• Targeted extraction: You know which tool to use and just want the model to extract the parameters

Mode comparison

Advanced strategies

Alternate modes in a workflow

A smart workflow can change tool_choice depending on the step:

# Step 1: force search
response1 = client.chat.completions.create(
    model="grok-4.5",
    messages=messages,
    tools=tools,
    tool_choice={"type": "function", "function": {"name": "search_products"}}
)

# Step 2: let the model decide (analysis or question)
response2 = client.chat.completions.create(
    model="grok-4.5",
    messages=updated_messages,
    tools=tools,
    tool_choice="auto"
)

Combine with the system prompt

The system prompt can guide the model in its tool choices even in auto mode:

messages = [
    {
        "role": "system",
        "content": "You are a travel assistant. Always use "
                   "get_weather before recommending outdoor activities."
    },
    {"role": "user", "content": "What should I do in Marseille this weekend?"}
]

Key takeaways

• auto: The model decides (default, recommended for most cases)
• required: Forces calling at least one tool
• none: Temporarily disables function calling
• Specific mode: Forces calling a specific tool by name
• Alternate modes in a workflow for fine control
• Combine tool_choice with the system prompt to guide the model