User Features

The features described in this section are accessible to any authenticated user of the application. AI Support offers the same conversations through two interfaces: the full Chat view, which provides a conversational interface for interacting with AI assistants, and the chat bubble, a floating window available on any page of the application. The bubble can also be shown on other websites as the embedded chat, where anonymous visitors talk to an assistant.

Chat

The Chat view is the central feature of the AI Support appjar. It allows users to have conversations with AI-powered assistants that can answer questions, provide guidance, and reference documentation relevant to the application. The view is organized as a two-panel layout: the main conversation area on the left and a session sidebar on the right.

On a phone the two panels are shown one at a time. The session list fills the screen, and tapping a session opens its conversation. The Chat sessions button at the top of the conversation returns to the list, and so does the phone's Back, even when the conversation was opened from a link or started as a new chat. The view keeps the panel the user chose when the phone is rotated, and returns to the list when the open session is deleted.

Starting a New Conversation

To begin a new conversation, click the + Chat button located at the top of the session sidebar. The main area displays To start a conversation, select an assistant and send a message. The assistant selector is located in the toolbar at the top of the conversation area, represented by a robot icon. Clicking this icon opens a dropdown menu listing all available assistants.

Once an assistant is selected, type a message in the input field at the bottom of the conversation area and click the Send button or press Enter. The AI assistant processes the message and streams a response in real time. While the response is being generated, the assistant is shown as is thinking, and a Stop button becomes available, allowing the user to halt generation at any point if the answer is already sufficient or headed in the wrong direction. A message whose generation was stopped early is marked with an Early badge and the note Generation stopped by followed by the name of the user who stopped it. If generation fails, the message is marked Failed. Administrators also get a private system message, The model could not answer, followed by the reason the provider gave, with any credentials removed.

The assistant assigned to a new conversation is the one marked as the default by an administrator. Users can change the assistant before or during the conversation using the assistant selector in the toolbar. If the assistant is changed mid-conversation, subsequent responses use the newly selected assistant's configuration, including its language model, prompt, and document categories.

An assistant can only be used when its own configuration, its language model, and its prompt are all enabled. Assistants that fail this test are still listed, and selecting View status on one opens the Assistant status dialog, which shows the assistant, its LLM, and its prompt as separate rows marked enabled or disabled. This explains why an assistant is unavailable without requiring administrator access.

Draft Messages

Text typed into the message input is preserved per assistant. When the user returns to a conversation with unsent text, a dialog offers to Restore draft or to Ignore it. Drafts are removed once the message is sent, and when a session is closed.

The Conversation Area

The conversation area displays all messages exchanged in the current session. Each message shows the sender's name and avatar, the timestamp, and the message content. Messages from the AI assistant are rendered with full markdown support, including code blocks with syntax highlighting, lists, headings, inline formatting, and Mermaid diagrams. A diagram that cannot be parsed is replaced by a notice that the diagram contains syntax errors.

Chat Conversation — An active conversation with an AI assistant

Hovering over an assistant message reveals a row of actions beneath it:

Copy message content: Copies the message text to the clipboard.
Inspect exchanged data: Opens the exchange behind the message in the LLM Inspector. Available to administrators.
Token count: The number of tokens the model used to generate the message, shown as a small badge.

Attachments on a message are listed below its text, each with a Download action, and a Preview action for the types that can be displayed in the browser.

Beside the message input sit its actions: the private mode switch, available to administrators and described in Private Mode, Attach files, and Record a voice note. Each appears only where it applies. When the input is too narrow to show them all, they collapse one by one into a More actions menu, where each item carries its label, such as Private mode: off.

Users can include the following in their messages:

Plain text: Standard text messages sent to the assistant. A message can be up to 2,000 characters long, unless the application sets another limit. While the input has focus, a counter such as 0/2000 shows how much of the limit is used, and the input stops accepting text when it is reached.
File attachments: Files can be attached to a message before sending. Click the paperclip icon (Attach files) beside the message input to open the attachment dialog. The dialog contains a file upload area where files can be selected from disk or dragged and dropped, and states the limits that apply: the maximum size per file (10 MB unless the administrator raises it), the maximum number of files per message (three by default), and the accepted file types. To remove a file before sending, click the remove icon on its chip in the attachment selector; Remove attachments discards all of them at once. Supported file types include documents (PDF, plain text, Markdown, CSV, DOC, DOCX, XLS, XLSX, LOG), images (JPG, JPEG, PNG), audio (MP3, WAV, OGG, OGA, OPUS, M4A, AAC, FLAC, AIF, AIFF), and video (MP4, AVI, WebM, MKV, MOV). Attached files are included as additional context alongside the message text. Documents and PDF files can always be attached. Images, audio, and video can be attached when the language model behind the selected assistant accepts them, and audio also when the application can transcribe it for the assistant. When the assistant changes, attachments it does not support are removed, with the notice Some attachments were removed because they are not supported by the current assistant. A message cannot be sent while its attachments are still uploading: the chat asks to Wait for the attachments to finish uploading before sending the message.
User mentions: In multi-participant conversations, users can mention other participants to direct their attention to a specific message. Typing @ opens the mention list, and the characters that follow narrow it by username. Each entry shows the person's name with the username it will type underneath, and a role badge such as Assistant, Admin, or User; the user's own entry is marked You. A mention that matches a user is highlighted in the message editor, and in the sent message it appears as the typed @username. The mentioned user is notified that they were mentioned, except for a private message, which only notifies administrators. Regular users can mention the active participants of the conversation. Administrators can also mention any other user of the application, which adds them to the conversation; when more users match than the list can show, it says More users match. Keep typing to narrow the search. Two mention targets are not users, and they lead the list: @assistant, shown with the assistant's name, asks the assistant for a response, which is how an answer is requested while the conversation is in manual mode, and @administrator, shown as Administrators, asks for human help.

The assistant does not always read an attachment in full. A long text file is cut, a long PDF file may be read only up to a number of pages or only as text, and a file that cannot be read is left out. When that happens, the user who attached the file, and only that user, sees a warning naming it, such as report.pdf is too long: the assistant read only its first 10 pages. or notes.docx could not be read.

Voice Notes

Where the chat offers attachments, users can also speak instead of typing, provided the assistant's language model hears audio or the application can transcribe it. Record a voice note (the microphone) starts recording, after the browser asks for permission to use the microphone. While recording, a bar shows the elapsed time with three buttons:

Discard: Throws the recording away.
Pause / Resume: Pauses the recording and continues it.
Send: Sends the voice note, together with any text typed in the input, as one message.

A voice note can be up to two minutes long. At that point recording pauses, and the elapsed time explains that The voice note reached its maximum length; the note can still be sent or discarded. When the microphone cannot be used, the chat says why, for example Allow the microphone to record a voice note or No microphone was found.

Audio attachments, voice notes included, play in place in the conversation, with a waveform. When the application can transcribe audio, Show transcript beneath the player displays what is said, showing Transcribing… the first time while the transcript is produced; Hide transcript folds it away again. A transcript is produced once and kept with the audio.

Session Management

Every conversation in AI Support is organized as a session. Sessions appear in the sidebar on the right side of the Chat view, listed in reverse chronological order based on their last activity. Each session card in the sidebar displays the session summary (automatically generated from the conversation content), the session number, the assistant name, the time of last activity, participant avatars, and the current rating if one has been provided.

The summary is generated right after the assistant's first answer and is then refreshed as the conversation continues, at most once every ten minutes by default. Only messages everybody in the session can see contribute to it; private and system messages never do.

A card also carries two indicators. A counter shows how many messages have arrived since the user last opened that session, as 1 new message or N new messages. An Alerts badge marks sessions the user has subscribed to.

Sessions have three possible statuses:

Active: The conversation is ongoing. Both the user and the assistant can exchange messages.
Archived: The conversation has been archived. It remains visible for reference but is no longer active for new messages.
Closed: The conversation has been explicitly closed. No further messages can be sent. When a non-administrator user closes a session, a rating dialog is presented automatically.

The actions menu on each session card (represented by a vertical three-dot icon) offers:

Close: Closes the session after a confirmation. A session cannot be closed while the AI assistant is generating a response.
Delete: Permanently removes the session with all its messages, attachments, draft messages, and participant information. A session cannot be deleted while the assistant is answering.
Link: Copies a link to the session to the clipboard, so a conversation can be shared with another user who has access to it. Individual messages have their own Copy message link action.
Subscribe / Unsubscribe: Turns alerts for the session on or off. See Alerts.

Warning

Deleting a session is a permanent action. All messages, attachments, and associated data will be permanently removed and cannot be recovered.

Alerts

By default a user is notified about a conversation only while looking at it. Subscribing to a session asks for a notification every time a new message is published in it, which is useful for a conversation that someone else is expected to answer.

Use Subscribe in the session actions menu and confirm; subscribed sessions show an Alerts badge in the sidebar. Unsubscribe stops the notifications. Subscriptions are per user, so subscribing does not affect what other participants receive.

Searching and Filtering Sessions

The session sidebar provides two tools for finding specific conversations.

Search opens the Search sessions or messages dialog, which matches the text against both session summaries and message content. When both kinds match, the results are split into two tabs that report how many of each were found, such as Sessions: 3 and Messages: 12; when only one kind matches, the results are shown directly, without tabs. A regular user never finds administrators' private messages. When a search matches more than the dialog can show, it says so and asks for a narrower term. A Show closed chats option includes closed sessions in the results. Selecting a result opens that session, positioned on the matching message.

Filters opens the Session filters dialog, whose fields can be combined:

Summary: Matches text in the session summary.
Status: Filters by active, archived, or closed.
Rating: Filters by the rating given to the session, with an Unrated option for sessions still awaiting one.
Created by: Filters by the user who started the session.
Participants: Filters by the users taking part, with Active participants only to exclude those who have left.
Assistant: Filters by the assistant assigned to the session.
Channel and Type: Filter by the messaging channel a conversation arrived through and by the session type.
Creation date ranges and Last activity date ranges: Filter on either timestamp, using From, Until, and At time bounds. Use last activity date switches which of the two timestamps the range applies to.

Participants

By default, a conversation involves only the user who created it and the AI assistant. Additional users can be invited to join. Use the Search users dropdown in the toolbar at the top of the conversation area to find and add participants; every action asks for confirmation first, naming the user affected. Once added, all participants can read and send messages within the session.

The participants panel lists everyone who has taken part, each by name with their username underneath when the two differ, and with the moment they joined or left. The user who started the conversation is marked Started this chat. A participant who has not written yet is shown as such rather than with an empty activity.

The available actions depend on who the participant is:

Remove takes another participant out of the conversation; Leave is the same action applied to oneself.
Bring back returns a participant who had been removed; Join back is the same action applied to oneself.
An administrator viewing a conversation they are not part of can join it directly.

The assistant itself is a participant and can be replaced by another one, which the confirmation describes as replacing one assistant with the other.

Every arrival and departure is recorded in the conversation as a system message, such as Olivia added Leo to the conversation or Leo left the conversation, so the history explains who could see what and from when. A user who is removed while viewing the session is told that they have been removed.

All participants see messages in real time. Typing indicators show when another participant is composing a message, naming the user, or reporting that multiple users are typing.

Rating a Conversation

After a session is closed, users are prompted to rate the conversation. The rating dialog allows the user to assign a numerical rating and optionally provide a comment explaining their experience. This feedback is visible to administrators and can be used to assess the quality of AI responses and fine-tune the assistant configuration.

Sessions that have been closed but not yet rated are highlighted in the sidebar, and a notification banner appears at the top of the Chat view reminding the user to rate their pending sessions. Users can filter the session list to show only unrated sessions for easy access.

Automatic and Manual Mode

Conversations can operate in two modes:

Automatic mode: This is the default mode. The AI assistant automatically processes each user message and generates a response. This is the standard conversational experience.
Manual mode: In this mode, the AI assistant does not automatically respond to user messages. This mode is useful when a human administrator or support agent takes over the conversation and handles responses directly.

The mode is toggled with the Switch assistant response button in the conversation toolbar. In the chat bubble, administrators switch it from the Chat actions menu instead, with Switch to manual or Switch to automatic. Switching to manual mode is useful when the AI response is not sufficient and human intervention is needed, while switching back to automatic mode resumes AI-assisted responses. Either change is recorded in the conversation as a system message stating that assistant response has been enabled or disabled, and by whom.

In manual mode a participant can still ask the assistant for a single answer by mentioning it, as described below.

Private Mode

Administrators can write messages that only other administrators see. The padlock action beside the message input, Switch admin-private conversation, turns private mode on and off for the messages that follow; in the More actions menu of a narrow input it reads Private mode: on or Private mode: off. Turning it on reports Private mode activated. Your messages are visible to administrators only, and turning it off reports that they are visible to everybody again.

Private messages are visually distinct in the conversation area and carry a padlock, so it is always clear which part of a thread the end user can read. They are useful for internal notes, coordination between support staff, or annotations that should not reach the customer. The LLM Inspector marks the same messages Private.

Private mode is only available to administrators. The assistant still answers a message sent in private mode, but its reply is private as well, so the exchange stays hidden from the other participants.

Private messages also stay away from what regular users can see indirectly. When the assistant answers a message everybody can see, the private messages of the conversation are not sent to the language model, so the answer cannot reveal them. They are left out of the session summary, and they never appear in a regular user's message search.

Private mode works the same way in the chat bubble. A new conversation always starts with private mode off, and an existing conversation keeps the state it was left in, whether it is opened in the Chat view or in the bubble.

Free Mode Limits

Without a licence, AI Support allows 50 messages per day across the application. A Daily Messaging badge at the top of the Chat view counts the messages sent today against that limit, and changes colour as the limit approaches. Its tooltip states when the count resets, expressed in the user's own time zone; the allowance resets at midnight.

Once the limit is reached, sending is refused with Messaging limit reached. Contact the administrator, and attachments report Free limit reached. Installing a licence removes the limit and the badge.

The Chat Bubble

Opening the Chat view means leaving whatever page the user was on. For a short question — what a field means, how to complete the task in front of them — that is often too much interruption. The chat bubble solves this: a floating button sits above the application on every page, and opens a compact chat window over the current page without navigating away.

The bubble is not a reduced copy of the Chat view. It works on the same conversations, with the same assistants and the same messages, so a conversation can be started in the bubble and continued later in the full view, or the other way round. Attachments, voice notes, mentions, markdown responses, streaming, private mode, and the free-mode limit all behave as described above.

Moving and Resizing

The floating button can be dragged anywhere on the page, which matters when it covers something the user needs to see. It cannot be dragged off-screen: released near an edge, it animates back to sit within a small margin of the viewport, and it is repositioned automatically if the browser window is resized. Dragging is distinguished from clicking by distance, so a small movement still counts as a click and opens the window.

The chat window itself can be resized by dragging any of its edges or corners, from a compact box up to a large panel, for conversations that need more room. Expand chat window in the header does the same in one click, switching between the standard size and a larger one; once expanded the same control becomes Restore chat window size. Reset bubble position returns the floating button to its default corner after it has been dragged somewhere inconvenient.

Both the button's position and the window's size apply to the current page visit. Reloading the page restores the defaults, unless the application is configured to remember where the button was dragged for a while.

On a narrow screen the window opens full screen instead, since there is no room for a floating panel. Resizing and repositioning make no sense in that mode, so Expand chat window and Reset bubble position are shown but disabled. On a phone, the phone's Back closes the window.

The Window Header

The header of the chat window has two rows. The upper row names the conversation currently loaded and carries the controls that act on the window:

Expand chat window / Restore chat window size: Switches the window between its standard and enlarged size.
Reset bubble position: Moves the floating button back to its default position.
Close window (×): Hides the window. The conversation is left untouched, and clicking the floating button opens it again.

The lower row carries the controls that act on conversations:

Search: Opens the search dialog, which looks through the user's conversations and their messages. Choosing a result loads that conversation and scrolls straight to the matching message.
Filter: Narrows the conversation selector using the same session filters as the Chat view. A badge indicates when filters are active.
The conversation selector: A dropdown listing the user's conversations, each with its summary, number, and assistant. A conversation with unread messages shows the count beside it. Selecting one loads it into the window; typing in the field searches the list by summary.
Chat actions: A menu holding the actions that apply to the loaded conversation — Link copies a deep link to it; Switch to manual or Switch to automatic changes its mode, shown only to administrators, on a conversation that is not closed and has an assistant; Close closes the conversation after a confirmation; and Delete removes it after a confirmation. Each entry is available only when the action applies, and the whole menu is disabled until a saved conversation is loaded.
New chat (+): Starts a new conversation, exactly as + Chat does in the Chat view.

Staying Informed While Working Elsewhere

The point of the bubble is that it can be ignored until it matters, so it reports activity in two ways.

The floating button carries a badge with the number of unread messages across all of the user's conversations, not only the one currently loaded. It is the single indicator to watch while working elsewhere in the application.

A message arriving in a conversation other than the one open in the window also raises a short notification showing a preview, which is the fastest way to notice a reply without watching the badge. These previews are deliberately limited: a conversation only raises them if the user has subscribed to its alerts, and a user is never notified of their own messages, of messages they have already seen, or — unless they are an administrator — of administrator-only messages. Streaming fragments and system messages do not raise notifications either, so a single reply produces one notification rather than a stream of them.

Being mentioned, and being added to or removed from a conversation, are also reported through the bubble, so an invitation to a conversation reaches the user wherever they are in the application.

The Embedded Chat

The chat bubble can also be placed on other websites, such as a public marketing site, so that people who are not users of the application can talk to an assistant. Setting it up is described in Setting up the Embedded Chat.

Visitors of such a site do not sign in. Each browser is recognised by a cookie, and its visitor appears in the application as a user named visitor- followed by a short code. A visitor finds their earlier conversations again on a later visit from the same browser, as long as the browser keeps the cookie. Administrators see these conversations in the Chat view and the LLM Inspector, as sessions of the site's web channel, and can join and answer them like any other.

For visitors, the bubble works as described in The Chat Bubble, with these differences:

The chat is shown in the language of the page that embeds it.
Attachments, mentions, the conversation filter, and links to conversations are turned off, unless the application enables them.
No pop-up notifications are shown on the site; the unread badge on the floating button carries the news instead.
A position the visitor drags the button to is remembered for a short while, so it stays in place as the visitor moves between the pages of the site.
Visitors can send only a limited number of messages in a given period. A visitor who reaches a limit loses the bubble: the message is not sent, and the bubble stays hidden, also on the pages loaded afterwards, until the limit's period has passed.