Tool call

NavTalk supports integration of digital humans with external systems through function calling, enabling functionalities such as weather inquiries, knowledge base retrieval, and business system triggers.

🔹 Step 1: Define Function Tools (tools)

When establishing a WebSocket session, configure the tools in the session.update request body:

tools: [
  {
    "type": "function",
    "name": "function_call_judge",
    "description": "Automatically triggers function extension calls when user requests exceed the current conversation capabilities.",
    "parameters": {
      "type": "object",
      "properties": {
        "userInput": {
          "type": "string",
          "description": "Raw user request content to be processed"
        }
      },
      "required": ["userInput"]
    }
  }
]

✅Note: No explicit user invocation is necessary; the AI will automatically trigger upon recognizing command intent in the conversation.

🔹 Step 2: Receive Callback Events

When the function is triggered, the WebSocket will receive messages of the following type:

if (data.type === "response.function_call_arguments.done") {
  handleFunctionCall(data);
}

field	meaning
type	Fixed as response.function_call_arguments.done
arguments	The parameters for the function call (in JSON string format)
call_id	The current unique identifier for the function call

🔹 Step 3: Handle Function Calls

function handleFunctionCall(eventJson) {
  try {
    const functionCallArgs = JSON.parse(eventJson.arguments);
    const userInput = functionCallArgs.userInput;
    const callId = eventJson.call_id;

    handleWithMemAgent(userInput)
      .then(result => {
        sendFunctionCallResult(result, callId);
      })
      .catch(error => {
        console.error("Function call failed: ", error);
      });

  } catch (error) {
    console.error("Parameter parsing failed: ", error);
  }
}

🔹 Step 4: Call Backend Interfaces to Handle Business Logic

You can pass the userInput to the business system or middleware interface, such as a weather API or CRM system:

function handleWithMemAgent(userInput) {
  return new Promise(async (resolve, reject) => {
    const chatId = await getFromChromeStorage("chatId");

    fetch(`${baseUrl}/api/realtime_function_call`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        userInput,
        license,
        chatId
      }),
    })
    .then(response => response.text())
    .then(result => resolve(result))
    .catch(error => reject(error));
  });
}

🔹 Step 5: Return Results to AI Conversation

After processing the results, they need to be pushed back to the WebSocket, prompting the AI to continue responding:

function sendFunctionCallResult(result, callId) {
  const resultJson = {
    "type": "conversation.item.create",
    "item": {
      "type": "function_call_output",
      "output": result,
      "call_id": callId
    }
  };

  socket.send(JSON.stringify(resultJson));

  // Actively trigger AI response
  socket.send(JSON.stringify({ "type": "response.create" }));
}

🎯 Common Application Scenario Examples

scene	user input	system action
Check the weather	"What will the weather be like in Beijing tomorrow?"	Call the weather API and return the weather forecast
Query the knowledge base	"How to integrate Function Calling with NavTalk?"	Access the internal knowledge system to retrieve answers
Control devices	"Turn on the conference room lights."	Call the IoT interface to control hardware devices