> ## Documentation Index
> Fetch the complete documentation index at: https://docs.e6data.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Assistant

> agent

ReferenceApplies to: Unified engine deployments

`[agent]`

## Settings and limits

### `enabled`

Chooses whether the built-in assistant is started.

* **Default:** `false`
* **Type:** boolean

> **Example**
>
> `false` keeps the assistant off, so its model settings are not needed. `true` also needs the agent role, HTTP access, and a valid model connection.

### `max_steps`

Limits how many rounds of model reasoning and tool calls one assistant turn can make.

* **Default:** `32`
* **Type:** integer; positive

> **Example**
>
> With `32`, one user request cannot continue through more than 32 model rounds. Lower it when you want to bound long chains of tool use.

### `max_output_tokens`

Limits how much text and other generated output the model may return in one response.

* **Default:** `8192`
* **Type:** integer; positive

> **Example**
>
> `8192` requests an output budget of 8192 model tokens. Tokens are pieces of text, not bytes or complete words.

### `turn_timeout_seconds`

Limits how long one assistant turn may run.

* **Default:** `900`
* **Type:** integer (seconds); positive

> **Example**
>
> `900` gives one turn up to 15 minutes, including model calls and tools. It is separate from a SQL task’s execution timeout.

### `max_input_bytes`

Limits the size of a user message together with the page information sent to the assistant.

* **Default:** `65536`
* **Type:** integer (bytes); positive

> **Example**
>
> `65536` allows up to 64 KiB of input. A large pasted error log counts toward this limit.

### `max_tool_result_bytes`

Limits how much of one tool response the assistant can retain for a turn.

* **Default:** `131072`
* **Type:** integer (bytes); positive

> **Example**
>
> `131072` gives one result up to 128 KiB. This bounds a returned catalog listing or query response before it becomes part of the turn.

### `max_turn_bytes`

Limits the total size of one saved assistant turn.

* **Default:** `2097152`
* **Type:** integer (bytes); positive

> **Example**
>
> `2097152` gives the saved turn a 2 MiB budget for its retained input, tool results, and other turn data.

### `max_page_size`

Limits how many records one list response may return.

* **Default:** `100`
* **Type:** integer; positive

> **Example**
>
> `100` returns a long list over pages of at most 100 items. A client can still request a smaller page.

### `poll_interval_seconds`

Sets the shortest wait between assistant checks of a running engine operation.

* **Default:** `5`
* **Type:** integer (seconds); positive

> **Example**
>
> `5` means it waits at least five seconds between status checks. Lower values notice progress sooner but make more requests.

## Model connection

### `model.provider`

Chooses the request and response format used to talk to the model service.

* **Default:** `"openai"`
* **Type:** choice: `openai`, `openai_responses`, `anthropic`

> **Example**
>
> Use `"openai_responses"` for a Responses API endpoint, `"openai"` for a Chat Completions endpoint, or `"anthropic"` for that provider format. Match the format to the configured service URL.

### `model.base_url`

Tells the assistant where to send model requests.

* **Default:** `""`
* **Type:** string

> **Example**
>
> A provider may give you an API root ending in `/v1`. Put that root here; Unified adds the path required by the selected provider format.

### `model.model`

Chooses the model requested from the configured model service.

* **Default:** `""`
* **Type:** string

> **Example**
>
> Use a model name accepted by your provider. Changing this field changes the requested model; it does not change the provider URL or API format.

### `model.credential`

Names the saved credential whose token the assistant uses for model requests.

* **Default:** Unset
* **Type:** string; optional

> **Example**
>
> If the saved bearer credential is named `assistant-model`, set this to `"assistant-model"`. The actual token is read when making a provider request.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.