Tools And Mid Call Function Calling

Operator declares a tool with name='check_balance' but no parameters JSON schema. Model emits tool calls with random argument shapes.

Define every tool with the documented OpenAI-style schema {name, description, parameters: {type:'object', properties, required}}. The model uses the JSON Schema to format arguments; without it the arg JSON is ad-hoc and breaks serverUrl handlers.

Tool 'send_followup_sms' is long-running and the operator does not want to block the assistant. They mark it async:true.

async:true makes Vapi not wait for a result before continuing the conversation. Operator's serverUrl still returns 2xx promptly and runs the work in background. Do not use async:true for tools whose result must be spoken — the assistant won't see the response.

serverUrl receives the tool-calls webhook with message.toolCalls[] containing id, type, function.name, function.arguments. Operator parses arguments as a Python dict.

function.arguments is a JSON-encoded string per the OpenAI tool-call shape. json.loads it before use. Validate fields against the declared parameters schema before executing side effects. Reply with results[] keyed by toolCallId.