Function Calling

Function Calling

Function calling (tool use) lets AI models invoke external functions. AI-Lib provides a unified interface for tool calling across all providers that support it.

Defining Tools

Rust

use ai_lib::ToolDefinition;
use serde_json::json;

let get_weather = ToolDefinition {
    name: "get_weather".into(),
    description: Some("Get current weather for a city".into()),
    parameters: json!({
        "type": "object",
        "properties": {
            "city": {
                "type": "string",
                "description": "City name"
            },
            "unit": {
                "type": "string",
                "enum": ["celsius", "fahrenheit"]
            }
        },
        "required": ["city"]
    }),
};

Python

get_weather = {
    "name": "get_weather",
    "description": "Get current weather for a city",
    "parameters": {
        "type": "object",
        "properties": {
            "city": {
                "type": "string",
                "description": "City name",
            },
            "unit": {
                "type": "string",
                "enum": ["celsius", "fahrenheit"],
            },
        },
        "required": ["city"],
    },
}

Non-Streaming Tool Calls

Rust

let response = client.chat()
    .user("What's the weather in Tokyo?")
    .tools(vec![get_weather])
    .execute()
    .await?;

for call in &response.tool_calls {
    println!("Function: {}", call.name);
    println!("Arguments: {}", call.arguments);
    // Execute the function and send results back
}

Python

response = await client.chat() \
    .user("What's the weather in Tokyo?") \
    .tools([get_weather]) \
    .execute()

for call in response.tool_calls:
    print(f"Function: {call.name}")
    print(f"Arguments: {call.arguments}")

Streaming Tool Calls

Tool calls stream as partial events that the pipeline’s Accumulator assembles:

Rust

let mut stream = client.chat()
    .user("What's the weather?")
    .tools(vec![get_weather])
    .stream()
    .execute_stream()
    .await?;

while let Some(event) = stream.next().await {
    match event? {
        StreamingEvent::ToolCallStarted { name, id, .. } => {
            println!("Starting tool: {name} (id: {id})");
        }
        StreamingEvent::PartialToolCall { arguments, .. } => {
            print!("{arguments}"); // Partial JSON arguments
        }
        StreamingEvent::ToolCallEnded { id, .. } => {
            println!("\nTool call {id} complete");
        }
        StreamingEvent::ContentDelta { text, .. } => {
            print!("{text}");
        }
        _ => {}
    }
}

Python

async for event in client.chat() \
    .user("What's the weather?") \
    .tools([get_weather]) \
    .stream():
    if event.is_tool_call_started:
        call = event.as_tool_call_started
        print(f"Starting: {call.name}")
    elif event.is_partial_tool_call:
        print(event.as_partial_tool_call.arguments, end="")
    elif event.is_content_delta:
        print(event.as_content_delta.text, end="")

How It Works

1. You define tools and pass them in the request
2. The protocol manifest maps tools to the provider-specific format
3. The model decides to call a tool (or respond with text)
4. For streaming, the pipeline’s Accumulator assembles partial tool call chunks
5. You receive unified ToolCallStarted, PartialToolCall, and ToolCallEnded events

Provider Support

Check the provider’s capabilities before using tools:

Provider	Tool Calling
OpenAI	Supported
Anthropic	Supported
Gemini	Supported
DeepSeek	Supported
Groq	Supported
Mistral	Supported
Qwen	Supported

The manifest’s capabilities.tools: true flag indicates support.