Add Tool

Add a Tool for Function Calling

Creates callable tools that LLMs can invoke during conversations via Prompture's tool use system.

Before Starting

Ask the user for:

• Tool function name and what it does
• Parameters — names, types, and which are required vs optional
• Return type — what the tool returns (string result fed back to the LLM)
• Registration method — decorator (@registry.tool) or explicit (registry.register())
• Target driver — which provider will use the tools (must have supports_tool_use = True)

Key Concepts

ToolDefinition

A dataclass describing a single callable tool:

from prompture.tools_schema import ToolDefinition

td = ToolDefinition(
    name="get_weather",
    description="Get current weather for a city.",
    parameters={
        "type": "object",
        "properties": {
            "city": {"type": "string", "description": "City name"},
            "units": {"type": "string", "description": "Temperature units"},
        },
        "required": ["city"],
    },
    function=get_weather_fn,
)

tool_from_function — Auto-generate from type hints

from prompture.tools_schema import tool_from_function

def get_weather(city: str, units: str = "celsius") -> str:
    """Get the current weather for a city."""
    return f"Weather in {city}: 22 {units}"

td = tool_from_function(get_weather)
# Automatically extracts: name, description (from docstring), parameters (from type hints)

Type mapping: str -> "string", int -> "integer", float -> "number", bool -> "boolean", list -> "array", dict -> "object". Optional[X] is unwrapped. list[X] becomes {"type": "array", "items": {...}}.

Required vs optional: Parameters without defaults are required. Parameters with defaults are optional.

ToolRegistry — Managing tool collections

from prompture import ToolRegistry

registry = ToolRegistry()

# Method 1: Decorator
@registry.tool
def search(query: str) -> str:
    """Search the knowledge base."""
    return f"Results for: {query}"

# Method 2: Explicit registration
def calculate(expression: str) -> str:
    """Evaluate a math expression."""
    return str(eval(expression))

registry.register(calculate)

# Method 3: With name/description override
registry.register(some_func, name="custom_name", description="Custom description")

# Method 4: Add pre-built ToolDefinition
registry.add(td)

Serialization Formats

# For OpenAI, Groq, Grok, Azure, Ollama (OpenAI-compatible format)
tools_list = registry.to_openai_format()
# Returns: [{"type": "function", "function": {"name": ..., "description": ..., "parameters": ...}}]

# For Anthropic/Claude
tools_list = registry.to_anthropic_format()
# Returns: [{"name": ..., "description": ..., "input_schema": ...}]

# Single tool
td.to_openai_format()
td.to_anthropic_format()

Conversation Integration

from prompture import Conversation, ToolRegistry

registry = ToolRegistry()

@registry.tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Sunny in {city}"

# Pass tools at construction
conv = Conversation(driver=driver, tools=registry)
result = conv.ask("What's the weather in Paris?")

# Or register after construction
conv = Conversation(driver=driver)
conv.register_tool(get_weather)

The conversation automatically:

1. Sends tools to the LLM via generate_messages_with_tools()
2. Parses tool_calls from the response
3. Executes the tool function with the provided arguments
4. Sends the result back to the LLM
5. Repeats until the LLM returns a final text response (or max_tool_rounds is hit)

Execution and Error Handling

# Direct execution
result = registry.execute("get_weather", {"city": "Paris"})

# Missing tool raises KeyError
registry.execute("nonexistent", {})  # KeyError: "Tool not registered: 'nonexistent'"

# Tool errors during conversation are caught and sent back as error strings
# The LLM sees: "Error: <exception message>" and can respond accordingly

Implementation Steps

1. Define the tool function

def my_tool(param1: str, param2: int = 10) -> str:
    """Clear, concise description — this becomes the LLM's tool description.

    The first line of the docstring is used as the description.
    """
    # Implementation
    return json.dumps({"result": "value"})

Rules:

• Always include type hints — they generate the JSON Schema
• Always include a docstring — first line becomes the tool description
• Return a string — this is what the LLM sees as the tool result
• For complex return data, return json.dumps(result)

2. Register in a ToolRegistry

registry = ToolRegistry()
registry.register(my_tool)

3. Use with a Conversation

conv = Conversation(
    driver=driver,          # Must have supports_tool_use = True
    tools=registry,
    max_tool_rounds=3,      # Safety limit (default: 3)
)

4. Write tests

def test_my_tool_schema():
    td = tool_from_function(my_tool)
    assert td.name == "my_tool"
    assert "param1" in td.parameters["properties"]
    assert td.parameters["required"] == ["param1"]

def test_my_tool_execution():
    reg = ToolRegistry()
    reg.register(my_tool)
    result = reg.execute("my_tool", {"param1": "test"})
    assert "result" in result

Drivers with Tool Use Support

The following drivers support generate_messages_with_tools():

• OpenAI (openai_driver.py, async_openai_driver.py)
• Claude (claude_driver.py, async_claude_driver.py)
• Groq (groq_driver.py, async_groq_driver.py)
• Grok/xAI (grok_driver.py, async_grok_driver.py)
• Azure (azure_driver.py, async_azure_driver.py)
• Ollama (ollama_driver.py, async_ollama_driver.py)
• Moonshot (moonshot_driver.py, async_moonshot_driver.py)
• OpenRouter (openrouter_driver.py, async_openrouter_driver.py)

Check driver.supports_tool_use before using tools with a driver.

Best Practices

• Keep tools focused — one function per action, not a kitchen-sink tool
• Use descriptive docstrings — the LLM relies on the description to decide when to call tools
• Type hints are schema — always annotate parameters; they generate the JSON Schema the LLM sees
• Return strings — tool results are injected into the conversation as text
• Handle errors gracefully — exceptions are caught and sent to the LLM as "Error: ..." messages
• Set max_tool_rounds — prevent infinite tool-calling loops
• Test schemas — use tool_from_function() + assertions to verify the generated schema matches expectations