Assistant chat
The in-app assistant explains Overvis HMI, reads the current workspace, inspects page elements you point at, and changes workspace data after your approval. It works with your permissions: live writes need the operator role or higher, and setup changes need editor, admin, or owner. It is available only in the signed-in cloud app, not in a local workspace. The steps for common tasks are in Get help from the assistant.
Opening the assistant
Section titled “Opening the assistant”The AI button in the header opens and closes the assistant panel over the current page. The
button shows how many chats have the assistant at work, and changes color when one needs you. On an
empty workspace, Set up your workspace with AI opens the same panel. A link with
?assistantChat=<chat id>, as in the assistant’s notifications, opens the panel on that chat.
Closing the panel, the tab, or the browser stops nothing: the assistant keeps working, and its cards wait until you answer them.
The panel
Section titled “The panel”| Part | What it does |
|---|---|
| Chat list | Start a new chat, open, rename, or delete a chat, and Stop any chat where the assistant works or waits. See Chats running at once |
| Chat header | Auto-allow and the monitor icon, Watch assistant preview. See Approvals and Auto-allow and previews |
| Transcript | A chat opens at its latest messages; older ones load as you scroll up. A reply shows the tools the assistant uses, and the view follows it while you are near the bottom |
| Live status | Thinking (tokens: N), Waiting for assistant with the elapsed time, or Summarizing earlier messages |
| Composer | The message box, the paperclip, the level menu, the element picker, the context ring, and Send. A draft with its attachments survives closing the panel |
| Queue | Messages written while the assistant works, above the message box. See Message queue |
| Approval card | Allow or Don’t allow a change. See Approvals and Auto-allow |
| Question card | Questions with answer buttons. See Question cards |
| Job wait card | The progress of a job the assistant started, such as reading a PDF or Excel file or searching for Modbus devices. Cancel stops it, and the assistant continues with what it has |
| Repeated-steps card | The assistant took the same step several times in a row. Continue lets it go on; Stop ends its work in the chat |
The empty chat says “Ask anything about Overvis HMI.” Every reply appears the same in every tab and on every device.
What the assistant sees and does
Section titled “What the assistant sees and does”The assistant answers in your language and searches this documentation. It sees what you have open: the selected rows of the parameter tree, the open report with its range and visible series, or the open dashboard with the status and active alarms it shows.
It can use the Overvis HMI API with your permissions, search the web and read web pages, open an app page for you, file a Helpdesk ticket or a problem report, and list your Helpdesk tickets. Each change to your data, each web page it reads, and each ticket or problem report needs your approval, unless you auto-allowed its category. Searching the web and opening a page need no approval.
Its replies link to the app pages they mention. After it creates or changes something, the reply links to each item: a dashboard or report to its viewing page, anything else to its Setup page. Links to app pages open in the same tab; other links open in a new tab.
It creates and changes automations and alarm rules, usually from a starter. It checks each one and tries automations with a dry run before you approve the save. Asked why one did not act, it reads the run history, run steps, and alarm log.
Levels
Section titled “Levels”The level menu after the paperclip sets the level for the next reply. Each level uses a different model.
| Level | Description |
|---|---|
| Free | Fast; suits simple questions |
| Standard | Balanced; the default |
| Pro | The most capable, for complex tasks |
A chat keeps its level, and you can change it at any message. A new chat starts on the level of your latest chat. A level chosen while the assistant works applies to the messages queued after it, not to the running reply. Continue uses the level shown.
Standard and Pro are charged to the workspace of the current page. On the pages of a suspended workspace, the menu shows them as unavailable and the assistant answers at Free; your chosen level applies again on other pages.
Cost and spending limit
Section titled “Cost and spending limit”What each reply costs is shown in the chat; the prices are in What is charged. Stopping a reply can still incur a charge for work already done.
The assistant has no step or time limit: a long task runs until it is done. Assistant spending limit in Account settings → Profile holds an amount. Each time the assistant’s work on one request has cost that much, it pauses and shows the cost so far: Continue lets it spend the same amount again, and Stop ends its work in the chat. The cost includes web searches and page fetches. The limit applies to all your chats; an empty field means no limit.
Approvals and Auto-allow
Section titled “Approvals and Auto-allow”Before a change, the assistant shows an approval card. The card names the affected objects, shows values before and after, lists every item of a bulk change, and states known consequences; destructive changes stand out. JSON details shows the exact request. If something relevant changes while the card waits, Allow does not run the old request, and the assistant prepares a new one.
Allow approves this action only. The arrow next to it offers Allow [category] for this chat, which approves the action and its whole category for the chat. Don’t allow declines it.
Auto-allow in the chat header lists the categories the assistant may run without asking: internet access, setup changes, working copies, historical data changes, and device writes. Each is chosen separately, applies only to the current chat, and stays chosen when you reopen it. Clearing a category brings the cards back for later actions. Choosing a category does not answer a card that is already waiting. Working copies covers the assistant’s own copies of workspace files, used to edit and test dashboards; publishing those edits to the workspace is a setup change.
Question cards
Section titled “Question cards”The assistant can ask questions with answer buttons. Each question takes one or several answers, as the card shows. Every question also has a text field for another answer or details. Submit answers sends them once every question has an answer. A Recommended badge marks the assistant’s suggestion without selecting it.
Open questions do not expire. They stay answerable after you reopen the panel or reload the page, although unsubmitted selections may be lost on reload. Submitted answers replace the form and cannot be edited. Stop cancels the question. Send now on a queued message also cancels it and gives the message to the assistant instead.
Questions only clarify a request: app forms still open in their own dialogs, and actions still need their approval.
Chats running at once
Section titled “Chats running at once”The assistant works on the server. After you send a message, it keeps working when the panel, the page, the tab, or the computer closes, and other chats can work meanwhile.
Up to three chats can have the assistant working or waiting for you at the same time; “Running chats: N of M” shows how many are in use. A chat counts until its reply finishes or is stopped, also while it waits for an approval or an answer. A message that would start one more chat is not sent, and says why.
Each chat in the list shows its state: Working, Needs you (an approval, a question, or a check after repeated steps or at the spending limit), or Stopped with an error. You can always stop or delete a chat.
A new chat is named after the assistant’s first reply, in the language you wrote in. The pencil next to a chat renames it; a name you give is never replaced.
When a chat finishes, fails, or needs you while no browser tab shows it, you get a notification that opens the panel on that chat. Steps that need the browser, such as opening a page, are skipped while no tab shows the chat.
Message queue
Section titled “Message queue”You can keep writing, attaching files, and picking elements while the assistant replies or waits. Enter puts the message in a queue above the message box. A queued message keeps the page it was written on as its context.
When the reply finishes, the assistant gets all queued messages together, in order. Don’t allow on an approval also hands them over. The queue belongs to the chat: it survives a reload and is the same on your other devices.
| Action | Effect |
|---|---|
| Send now | Gives the message to the assistant after its current step. On a waiting approval, it declines the action. On an open question, it cancels the question. On a job wait, it stops the job |
| Edit | Moves the message, with its attachments and picked elements, back into the message box. The queue shows it as Being edited; Put back in the queue returns it unchanged |
| Remove | Deletes it |
After Stop or an error, queued messages are not sent automatically. Each shows Send, which sends it and the messages before it as a new request. A message that could not be sent, for example because the connection was lost or too many chats are running, stays as Not sent with the reason, and Send, Edit, and Remove.
Long chats
Section titled “Long chats”The context ring beside Send shows how close the chat is to the point where the assistant summarizes older messages; its tooltip gives the percentage. Once it is full, the next message summarizes the older part first, and the panel shows Summarizing earlier messages. The transcript you see does not change. Switching to a level that holds less can fill the ring at once. A reply that reaches the level’s limit stops; your next message continues from the summary.
When a reply fails
Section titled “When a reply fails”The chat says why a reply failed:
| Message says | What to do |
|---|---|
| The model provider declined the request | Send it again, or choose another level and send it again |
| The model provider is overloaded | Try again in a few minutes, or choose another level |
| The assistant is unavailable on our side | Trying again does not help; contact support@overvis.com |
| The workspace on this page is suspended | Use the Free level, continue from another page, or ask the owner to check its billing account |
Interrupted replies
Section titled “Interrupted replies”When the connection to the model fails for a while, the assistant tries again on its own; text and actions from completed steps remain. A reply that still cannot finish is marked “The response was interrupted. The work may be incomplete.”, with Continue when it can resume. This also happens when you sign out, your session ends, or your access to the workspace changes; Continue from a session with access resumes it. Continue stays available after you reopen the chat and does not send your draft.
Stop ends the assistant’s work in the chat. It cancels pending approvals and does not undo completed actions.
Rewind
Section titled “Rewind”Rewind to here, under each of your messages (on hover, and always on touch screens), takes the chat back to before that message. The message and everything after it leave the chat, and the message returns to the message box with its attachments and picked elements, to edit and send again. Stop the assistant first; rewinding is unavailable while it works or waits in the chat. Queued messages stay in the queue.
The removed messages are kept for chat analysis, but the chat no longer shows them. Rewinding does not undo device writes, sent notifications, charges, or edits in the chat’s working copy.
When the assistant changed workspace setup after the message, the dialog offers Also revert workspace setup (off by default). Checked, it lists each workspace with the checkpoint before the assistant’s first change and previews everything the revert undoes, including other people’s changes since then. Reverting needs the editor, admin, or owner role; restoring the workspace name or logo needs admin or owner. Each revert can be undone in Activity history; see Revert workspace setup. Automatic checkpoints are kept for 30 days, so setup changes older than that cannot be reverted this way.
Working copies and dashboard previews
Section titled “Working copies and dashboard previews”For dashboard work, the assistant makes a working copy of the workspace files for the chat, where it edits, builds, and previews dashboards. It asks before it makes one, unless Working copies is auto-allowed, and again before it applies the changes to the workspace, unless Setup changes is auto-allowed. The time a working copy runs is charged as Assistant sandbox. A working copy unused for an hour is removed and made again when the chat needs it. While an approval to create or apply a working copy waits, the working copy is kept and its time is still charged.
While it prepares dashboard changes, the chat shows the current stage, such as check, build, preview, or publication, and its elapsed time.
While a reply uses the preview, a small live view of it sits above the message box with what the assistant is doing there, for example Clicking “Save”…. Selecting it, a browser step in the chat, or Watch assistant preview in the chat header shows the preview beside the chat; on a phone it fills the screen until Back to chat. The minimize button or Esc returns to the chat.
The preview is labelled Draft, not published: it shows the chat’s working copy, never the published workspace, and you can watch it but not control it. Its top bar shows the page path, whether it runs the dashboard’s source (Live code) or the bundle that publishing would upload (Built bundle), and the viewport size. Out of date means the assistant edited the code after it opened the page.
Attachments and page elements
Section titled “Attachments and page elements”| What | Rules |
|---|---|
| Files per message | Up to 10, of which up to 5 JPEG, PNG, or WebP images the assistant looks at. Attach with the paperclip, paste, or drop. A message can consist of attachments only |
| Documents | Only in a workspace where you are an editor, admin, or owner. Sending one adds it, with the message’s images, to Setup → Files. ZIP files unpack into a folder there |
| Images only | Stay in the chat. Without edit access, or on a page without a workspace, only images can be attached |
| Sent images | Show as thumbnails. A click opens a preview in the panel with previous, next, and download |
| Files for you | Files the assistant prepares for download are in Setup → Files. Programs for an MC252 are saved as edge programs |
To point at something on the page, use the pointer button beside the message box (“Insert any
element on the page into the chat”). In pick mode the page freezes and the element under the
pointer gets an outline. A click inserts a numbered token such as [1: Start] into the message;
Shift+click also attaches a screenshot of the element. Esc, the button, or sending leaves pick
mode. You can pick dashboard controls, Setup rows, report comments, open form dialogs, and pipes in
a drawing; picking works while a Setup dialog is open.
Each picked element also appears as a chip under the message box and is sent with the message, even if you delete its token from the text.
Chat recording and Chat ID
Section titled “Chat recording and Chat ID”The top of every chat says that chats are recorded and stored for automatic analysis and improvement of the assistant. The chat’s Chat ID, with a copy button, is below that notice. Support uses the Chat ID to find a conversation.