passive-yellow•4w ago

Tool Calls in Callback

I am currently using a callback to generate my assistant. Someone calls in, we look up the number in our backend, get the data we need to serve to the assistant and generate it dynamically. I want to add a tool call for this but since we are generating this all dynamically I'd prefer not to build a static tool call for each case we have to deal with, is there a way to serve the tool call in the dynamic assistant?

5 Replies

Vapi•4w ago

Yes, you can serve tool calls dynamically by returning a complete assistant configuration—including a dynamically generated tools array—in your callback response. Here’s a concise example:

res.json({
  assistant: {
    name: "Dynamic Assistant",
    model: {
      provider: "openai",
      model: "gpt-4o",
      tools: [
        {
          type: "function",
          function: {
            name: "processOrder",
            description: "Process an order dynamically",
            parameters: { /* your schema here */ }
          },
          server: { url: "https://your-server.com/tools/processOrder" }
        }
        // Add more tools as needed based on your backend logic
      ]
    }
  }
});

res.json({
  assistant: {
    name: "Dynamic Assistant",
    model: {
      provider: "openai",
      model: "gpt-4o",
      tools: [
        {
          type: "function",
          function: {
            name: "processOrder",
            description: "Process an order dynamically",
            parameters: { /* your schema here */ }
          },
          server: { url: "https://your-server.com/tools/processOrder" }
        }
        // Add more tools as needed based on your backend logic
      ]
    }
  }
});

Just build the tools array based on your backend data before responding to the assistant-request webhook—no need for static tool definitions.
Reference: See the "Return a Complete Assistant Configuration" section in the Vapi documentation. Sources: - Personalization with user information - Custom LLM Tool Calling Integration

⤵️ Mention @Vapi Support Bot for follow-ups 🐥

conscious-sapphire•4w ago

You don’t need to hardcode a separate tool for every case — you can pass dynamic tool definitions when generating the assistant so the right schema is available at runtime. I can help you set up your backend so it returns the data and the tool structure together, which VAPI will use in the session. This way, your assistant adapts based on the caller without extra static configs. Do you already have your backend returning JSON schema or should I show you how to wire that in? @Nemnel

passive-yellowOP•4w ago

ahh that's awesome thank you @Xcel we have our backend set up using json thanks, like I said in the initial prompt we are already dynamically generating each assistant

conscious-sapphire•4w ago

Since you already have JSON and dynamic assistants set up, the next step is wiring those tool schemas into VAPI so the assistant can call them cleanly at runtime. The best way to approach it is by letting your backend return both the assistant data and the tool definition, then injecting that schema directly into the session. From there, we can add validation and error handling so the tools scale across different cases without extra static configs. If you’d like, I can walk you through this privately and help optimize your flow so it’s production-ready. @Nemnel

passive-yellowOP•2w ago

@LordXcel I'm good

Tool Calls in Callback

Did you find this page helpful?