Response format

Output types

output is an array of items in the order the assistant acted. A typical reply: a few tool_call items in a row, then a final text.

Each output item is discriminated by its type field. The order matches what the assistant did.

real-world output

[
  {
    "type": "tool_call",
    "tool_call": {
      "name": "get_current_weather",
      "input": { "latitude": 55.75, "longitude": 37.62 },
      "output": { "type": "content", "content": [{ "type": "text", "text": "{\"temp\":12}" }] },
      "cost": 0
    }
  },
  { "type": "text", "text": "It's +12°C in Moscow right now…" }
]

text

{ "type": "text", "text": "Here is my answer…" }

file

Appears when the assistant generated a file (an image, for example).

{
  "type": "file",
  "file": { "id": "88", "name": "generated_image.png", "mime_type": "image/png" }
}

Build the download link from the storage base URL.

tool_call

A tool call together with its result. On success:

{
  "type": "tool_call",
  "tool_call": {
    "name": "get_current_weather",
    "input": { "latitude": 55.75, "longitude": 37.62 },
    "output": { "type": "content", "content": [{ "type": "text", "text": "…" }] },
    "cost": 0
  }
}

When the tool fails, output.type becomes error:

"output": {
  "type": "error",
  "error": { "message": "Request timeout", "reason": "fetch_timeout" }
}

What specific tools return is covered in Tool responses.

search

Appears when the provider is Perplexity (web search is built into the model).

{
  "type": "search",
  "search": {
    "results": [{ "title": "…", "url": "…", "snippet": "…", "source": "…" }],
    "citations": ["https://…"],
    "images": [{ "title": "…", "image_url": "…", "origin_url": "…" }]
  }
}

custom

Provider-specific data.

{ "type": "custom", "custom": { "from": "perplexity", "content": { … } } }

Thinking doesn't end up in output

Extended thinking blocks appear only in messages as { "type": "thinking", … }, never in output.