1 min read
Tool Choice Control: Fine-Tuning When and How AI Uses Tools
I wrote “Tool Choice Control: Fine-Tuning When and How AI Uses Tools” to share practical, production-minded guidance on this topic.
Tool Choice Options
from openai import OpenAI
import json
client = OpenAI()
tools = [
{
"type": "function",
"function": {
"name": "search_database",
"description": "Search the product database",
"parameters": {
"type": "object",
"properties": {
"query": {"type": "string"}
},
"required": ["query"]
}
}
},
{
"type": "function",
"function": {
"name": "create_order",
"description": "Create a new order",
"parameters": {
"type": "object",
"properties": {
"product_id": {"type": "string"},
"quantity": {"type": "integer"}
},
"required": ["product_id", "quantity"]
}
}
}
]
# Option 1: "auto" - Model decides (default)
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": "Hello!"}],
tools=tools,
tool_choice="auto" # Model may or may not use tools
)
# Option 2: "none" - Never use tools
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": "Search for laptops"}],
tools=tools,
tool_choice="none" # Model will respond without tools
)
# Option 3: "required" - Must use a tool
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": "What products do you have?"}],
tools=tools,
tool_choice="required" # Model must call at least one tool
)
# Option 4: Specific function - Force a particular tool
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": "I want to buy something"}],
tools=tools,
tool_choice={
"type": "function",
"function": {"name": "search_database"}
}
)
Dynamic Tool Selection
from typing import List, Optional, Literal
from enum import Enum
class ConversationState(Enum):
BROWSING = "browsing"
SELECTING = "selecting"
CHECKOUT = "checkout"
SUPPORT = "support"
class DynamicToolSelector:
"""Select appropriate tools and control based on conversation state"""
def __init__(self):
self.all_tools = self._define_all_tools()
def _define_all_tools(self) -> dict:
return {
"search_database": {
"type": "function",
"function": {
"name": "search_database",
"description": "Search products",
"parameters": {
"type": "object",
"properties": {"query": {"type": "string"}},
"required": ["query"]
}
}
},
"get_product_details": {
"type": "function",
"function": {
"name": "get_product_details",
"description": "Get product details",
"parameters": {
"type": "object",
"properties": {"product_id": {"type": "string"}},
"required": ["product_id"]
}
}
},
"add_to_cart": {
"type": "function",
"function": {
"name": "add_to_cart",
"description": "Add item to cart",
"parameters": {
"type": "object",
"properties": {
"product_id": {"type": "string"},
"quantity": {"type": "integer"}
},
"required": ["product_id", "quantity"]
}
}
},
"process_payment": {
"type": "function",
"function": {
"name": "process_payment",
"description": "Process payment",
"parameters": {
"type": "object",
"properties": {
"amount": {"type": "number"},
"payment_method": {"type": "string"}
},
"required": ["amount", "payment_method"]
}
}
},
"contact_support": {
"type": "function",
"function": {
"name": "contact_support",
"description": "Contact human support",
"parameters": {
"type": "object",
"properties": {
"issue": {"type": "string"},
"priority": {"type": "string"}
},
"required": ["issue"]
}
}
}
}
def get_tools_for_state(self, state: ConversationState) -> tuple:
"""Return (tools, tool_choice) based on state"""
if state == ConversationState.BROWSING:
tools = [
self.all_tools["search_database"],
self.all_tools["get_product_details"]
]
return tools, "auto"
elif state == ConversationState.SELECTING:
tools = [
self.all_tools["get_product_details"],
self.all_tools["add_to_cart"]
]
return tools, "auto"
elif state == ConversationState.CHECKOUT:
tools = [
self.all_tools["process_payment"]
]
# Force payment processing during checkout
return tools, {"type": "function", "function": {"name": "process_payment"}}
elif state == ConversationState.SUPPORT:
tools = [
self.all_tools["contact_support"]
]
return tools, "required" # Must escalate
return [], "none"
# Usage
selector = DynamicToolSelector()
def chat_with_state(message: str, state: ConversationState) -> str:
tools, tool_choice = selector.get_tools_for_state(state)
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": message}],
tools=tools if tools else None,
tool_choice=tool_choice if tools else None
)
return response
Safety Controls
class ToolSafetyController:
"""Control tool usage based on safety policies"""
def __init__(self):
self.dangerous_tools = {"delete_data", "process_payment", "send_email"}
self.require_confirmation = {"create_order", "update_profile"}
self.rate_limits = {}
def filter_tools(self, tools: List[dict], user_permissions: List[str]) -> List[dict]:
"""Filter tools based on user permissions"""
return [
tool for tool in tools
if self._user_can_use(tool["function"]["name"], user_permissions)
]
def _user_can_use(self, tool_name: str, permissions: List[str]) -> bool:
"""Check if user has permission for this tool"""
if tool_name in self.dangerous_tools:
return f"admin:{tool_name}" in permissions
return True
def get_tool_choice(self, tool_name: str, already_confirmed: bool) -> dict:
"""Determine appropriate tool choice with safety checks"""
if tool_name in self.dangerous_tools:
if not already_confirmed:
return "none" # Require explicit confirmation
if tool_name in self.require_confirmation:
return {
"type": "function",
"function": {"name": f"confirm_{tool_name}"}
}
return {"type": "function", "function": {"name": tool_name}}
def check_rate_limit(self, tool_name: str, user_id: str) -> bool:
"""Check if tool usage is within rate limits"""
key = f"{user_id}:{tool_name}"
# Implementation depends on your rate limiting system
return True
# Confirmation flow
def execute_with_confirmation(message: str, tools: List[dict],
pending_action: dict = None) -> dict:
"""Execute tools with confirmation for sensitive actions"""
if pending_action:
# User is confirming a pending action
if "yes" in message.lower() or "confirm" in message.lower():
# Execute the pending action
result = execute_tool(pending_action["tool_call"])
return {"status": "executed", "result": result}
else:
return {"status": "cancelled"}
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": message}],
tools=tools,
tool_choice="auto"
)
message = response.choices[0].message
if message.tool_calls:
call = message.tool_calls[0]
# Check if this requires confirmation
if call.function.name in ["process_payment", "delete_data"]:
return {
"status": "pending_confirmation",
"action": {
"name": call.function.name,
"args": json.loads(call.function.arguments),
"tool_call": call
},
"message": f"Are you sure you want to {call.function.name}?"
}
# Execute directly
result = execute_tool(call)
return {"status": "executed", "result": result}
return {"status": "response", "content": message.content}
Workflow Control
class WorkflowController:
"""Control tool usage through a defined workflow"""
def __init__(self):
self.workflow = {
"start": {"allowed_tools": ["search_database"], "next": "found_results"},
"found_results": {"allowed_tools": ["get_details", "search_database"], "next": "viewing_product"},
"viewing_product": {"allowed_tools": ["add_to_cart", "search_database"], "next": "cart_updated"},
"cart_updated": {"allowed_tools": ["checkout", "continue_shopping"], "next": "checkout"},
"checkout": {"allowed_tools": ["process_payment"], "next": "complete"}
}
self.current_state = "start"
def get_allowed_tools(self, all_tools: List[dict]) -> List[dict]:
"""Get tools allowed in current state"""
allowed_names = self.workflow[self.current_state]["allowed_tools"]
return [t for t in all_tools if t["function"]["name"] in allowed_names]
def transition(self, executed_tool: str):
"""Move to next state based on executed tool"""
next_state = self.workflow[self.current_state].get("next")
if next_state:
self.current_state = next_state
def can_proceed(self) -> bool:
"""Check if workflow allows proceeding"""
return self.current_state != "complete"
# Usage
workflow = WorkflowController()
while workflow.can_proceed():
allowed = workflow.get_allowed_tools(all_tools)
response = client.chat.completions.create(
model="gpt-4o-2024-08-06",
messages=[{"role": "user", "content": user_input}],
tools=allowed,
tool_choice="required" if len(allowed) == 1 else "auto"
)
if response.choices[0].message.tool_calls:
tool_name = response.choices[0].message.tool_calls[0].function.name
workflow.transition(tool_name)
Tool choice control is essential for building AI applications that are predictable, safe, and aligned with your business logic. Use these patterns to guide model behavior exactly as needed.\n\n## Takeaways\n\nAdd a concise, personal takeaway and recommended next steps here.\n