Skip to main content

Concepts

A quick tour of how Talk2View fits together.

Partner, end-user, app​

  • A partner is you — the software company integrating Talk2View. You get one partner key.
  • An end-user is one of your users, chatting inside your app with their Talk2View account (or as a guest, if you enable it).
  • Conversations are scoped to (partner, end-user), so the same person is isolated across different partner apps.

Tools and the interrupt/resume loop​

Tools are what make the assistant do things instead of only talking. You define a tool with a name, a description, a parameter schema, and an execute function:

{
  name: 'set_filter',
  description: 'Filter the current view by status.',
  parameters: {
    type: 'object',
    properties: {
      status: { type: 'string', description: 'Which items to show.', enum: ['open', 'closed'] },
    },
    required: ['status'],
  },
  execute: async (args) => {
    // runs in your app
    return 'Filtered to ' + String(args.status) + '.';
  },
}

When the assistant decides to call a tool, the agent pauses, the call is sent to your client, your execute runs locally, your result is sent back, and the agent resumes. The tools you register always run in your app, through this pause and resume, so inside your app the assistant can only do what your tools let it do.

Built-in tools​

Talk2View also has tools of its own, which run on its servers rather than in your app:

  • Skills and planning — on by default. The assistant can look up and load skills (instructions for particular tasks, including the ones you write on the dashboard's Skills page) and keep a to-do list for multi-step work.
  • Web search, maps and image generation — only if you turn them on in the dashboard. Maps and image generation are for signed-in users, not guests. Image generation also needs a client that says it can show images; @talk2view/sdk 0.21.3 doesn't yet, so it isn't available through the SDK.
  • Scheduling and email — these work only after an end-user connects their own scheduling or email account for your app, and the SDK can't connect one yet. So you can't offer them through the SDK, even if you turn them on in the dashboard.

Models and configuration​

The AI workflow itself is Talk2View's — you don't rebuild it. What you control from the dashboard: which model to use, a system prompt, which tools and skills are available, allowed websites (for popup sign-in), guest access, voice, and web search.

Streaming​

Replies stream back as events (text chunks, tool calls, completion). The packaged chat renders them for you. If you're building your own UI, let the SDK read the stream: the Talk2View client's chat() method yields typed events (see the ChatEvent type exported from @talk2view/sdk). The API reference lists the HTTP endpoints, but it doesn't describe the events in the stream.