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

# Ask AI

> The assistant in the Veri dashboard: it explains what is on screen and takes actions only with your permission.

Ask AI is the assistant in the Veri dashboard, for every signed-in user. Open it with the **Ask AI** button at the bottom right, with `⌘J` (`Ctrl+J`), or from the command palette. Ask about the page you are on and it answers from what is there: the training job, the deployment list, the dataset. When an answer calls for an action, it proposes one and waits for you to allow it.

Veri runs the model behind it, on a Veri deployment, and you are not billed for your questions.

Ask AI is only available inside the dashboard. There is no package or API key flow for adding it to your own site.

## What it sees

Each time you ask, the panel reads the current page's route and title, up to 8,000 characters of its visible text, and what is running and billing across your account.

Page content goes to the model as untrusted input. The assistant cannot change its own permission mode or approve an action because of something it read on a page: both come only from what you click.

## What it can do

Its tools run in your browser as you, so it can't see or change more than you can.

| Kind | Tools |
| - | - |
| Reads | List training jobs, get one job, read a job's metrics, list deployments, get one deployment, read a deployment's metrics, read your balance, estimate a cost, list datasets |
| Writes | Cancel a training job, stop a deployment, register a model |

Reads run as soon as the model asks, and show as a row in the thread. A write shows as a card with **Allow once**, **Deny** and **Edit**.

It can't start a training job: that's done from the [CLI, the API or an agent](https://docs.veri.studio/training.md#starting-a-job).

## Permission modes

The chip under the message box picks the mode, and `Shift+Tab` cycles through them. You start in Manual mode, and your browser remembers the choice.

| Mode | What runs without asking |
| - | - |
| **Manual mode** | Reads only. Every write waits for you. |
| **Ask mode** | Reads, and writes marked as safe. None of the three writes above is marked as safe today, so they still wait for you. |
| **Auto mode** | Everything, including writes. Each call still shows in the thread. |

> [!WARNING]
>
> In Auto mode, a write runs with no card to stop it, and none of the three has an undo. [Stopping a deployment](https://docs.veri.studio/deployments/api.md#post-v1deploymentsidstop) is final.

## Feedback and versions

Thumbs up or down under an answer, with an optional comment, records a score on that answer. A second rating of the same answer replaces the first. The chip in the panel's header names the version that answered.

**Settings, Ask AI** shows the versions, their thumbs-up rates and your recent ratings. The commands under **Train the next version** on that page are Veri's own operations tooling: they aren't part of the published CLI.

## Your chats

An Ask AI chat is stored as [request traces](https://docs.veri.studio/deployments/observability.md#request-traces) on Veri's deployment, and expires on the same retention schedule. A stored chat holds your questions, the replies, and the page text and tool results sent with them, which can include job configs, metrics and dataset names.

Veri copies a chat into a dataset to improve Ask AI only if the person who asked has said yes:

* No answer counts as no.
* A yes covers chats that start after it, not earlier ones.
* A no takes effect at once, and also stops earlier chats that have not been used yet.
* Members of the Veri team count as yes until they say no.

The panel doesn't ask yet. To read or set your answer, call the API with your key:

```bash theme={null}
curl "https://api.veri.studio/v1/helper/consent" \
  -H "Authorization: Bearer $VERI_API_KEY"

curl -X PUT "https://api.veri.studio/v1/helper/consent" \
  -H "Authorization: Bearer $VERI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"choice": "no", "wording_version": "docs"}'
```

`choice` is `yes` or `no`, and it is always your own answer. `wording_version` records the text you answered from, 1 to 64 characters of `A-Z a-z 0-9 . _ : -`. `GET` returns `choice`, `wording_version` and `decided_at`, each `null` until you answer. Every answer is kept in a history.


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