Response format

Tool responses

Built-in tools (weather, search, documents and so on) are enabled on the assistant in Aist Backend, not through the Agent API. This page shows what their responses look like in output.

Scope

The set of tools and their parameters live in the assistant settings (Aist Backend) and can’t be controlled through the Agent API. The client doesn’t enable these tools in the request. What matters to you is being able to read their results — they arrive in output[].tool_call.output. Tools that the client controls are client tools.

A successful result of any tool has the shape { "type": "content", "content": [{ "type": "text", "text": "…" }] }. The text usually holds a JSON string or plain text. Below is what each tool returns.

Data and utilities

Tool	`text` in the response
`get_time`	One line per time zone: Moscow and UTC, preceded by the user’s local time (when their time zone is known) and the zone the model asked about. The weekday is in Russian: `"Moscow (UTC+3): 2026-04-23 14:32:05, четверг (GMT+03:00)\nUTC: 2026-04-23T11:32:05Z"`
`get_current_weather`	OpenWeatherMap JSON: `{"weather":[…],"main":{"temp":12.3,…},"wind":{…}}`
`get_currency_info`	`{"base":"USD","target":"RUB","rate":92.45}`
`web_search`	Text with Perplexity search results.
`look_image`	A text description of the image (vision model).

Image generation

generate_image returns a file — both inside tool_call.output and as a separate file item in output[]:

"output": {
  "type": "content",
  "content": [{ "type": "file", "file": { "id": "88", "name": "generated_….png", "mime_type": "image/png" } }]
}

Native image models (e.g. google/gemini-2.5-flash-image) generate the picture without a tool_call — straight as a file item in output[] (a mixed “text + file” reply is possible too).

Documents

Tool	`text` in the response
`read_document`	`{"page":1,"total_pages":12,"elements":[{"type":"heading","text":"…"},…]}`
`get_document_content`	`{"offset":0,"limit":100,"total":247,"has_more":true,"elements":[…]}`
`read_table`	CSV: `date,name,amount\n2026-04-01,Smith,1500\n…`
`search_document_vector`	`{"results":[{"score":0.94,"page":8,"text":"…"},…]}`
`search_document_fulltext`	`{"results":[{"page":3,"element_type":"paragraph","text":"…"},…]}` (up to 10)

Collections (knowledge bases)

Tool	`text` in the response
`search_collection_vector`	`{"results":[{"score":0.97,"document":"refund-policy.pdf","page":1,"text":"…"},…]}`
`search_collection_fulltext`	`{"results":[{"document":"…","page":2,"type":"paragraph","text":"…"},…]}` (up to 10)

External tools (external_tools)

An assistant can call tools with their own callback URL (configured in Aist Backend). The model sees them under a name prefixed with ext_. In output they look like a regular tool_call — with whatever the external service returned:

{
  "type": "tool_call",
  "tool_call": {
    "name": "ext_book_room",
    "input": { "room": "#3", "time": "15:00" },
    "output": { "type": "content", "content": [{ "type": "text", "text": "Booking BK-2091" }] },
    "cost": 0
  }
}

How this differs from client tools: external_tools are called by the server (at the configured URL), and the task isn’t interrupted. client_tools interrupt the task and hand control to you — see Client tools.