Clicking the arrow on the right of the input field sends your prompt. While the model is answering, the same button turns into a square icon labeled Cancel - clicking it stops the running generation right away. The text already generated stays as a message; the model simply stops writing further.
How to tell whether something is still running. Before any output is visible, a time indicator with a seconds counter runs at the top of the emerging message; the text beside it moves from "Thinking" through "Still thinking" to "Hmm, this one takes a bit longer". As soon as the first text appears, the counter stops and is replaced by the fixed "Thought for X s" line - even if the model keeps writing afterward. To tell whether anything is still running overall, the send button is the reliable signal: as long as it reads Cancel, the model is still generating; once it turns back into the arrow, the response is complete. The "Thought for X s" line stays visible even after a page reload.
The pencil icon next to one of your own messages opens Edit Message; the regenerate icon under a model's response regenerates that answer (Regenerate). Both work anywhere in the chat — including in the middle of the conversation. Editing is unavailable for your own messages that contain @-mentions, and while your workspace's subscription is inactive.
Attention
If you edit or regenerate a message that is not the last one in the chat, every message that came after it disappears from the visible history — even if those messages covered completely different topics. There is currently no way to switch back to the earlier version (no version history, no "1/2" navigation). Before continuing from an earlier point in the chat, copy out anything important from the later messages, or start a new chat instead if you want to keep the full history intact.
Per chat, you choose the AI model that should answer the request. Depending on your workspace configuration, models from OpenAI, Anthropic, Google, Mistral and others are available. An EU flag next to a model indicates that it is hosted in an EU region and is GDPR-compliant. Reasoning models (with "Thinking" or "Reasoning" in the name) deliver more thorough results for complex tasks but take longer and consume more credits.
More on the model selection and which model fits when: AI Models.
Upload files or images – adds a file or image directly to the chat. The content is sent along to the model so it can read, summarize or analyze it.
Note
You can attach a maximum of 50 files per chat (documents and images combined). This limit applies the same way across all workspace types - it currently can't be increased individually, including under an Enterprise contract.
Note
For files such as PDF, Word, and txt, the text is extracted, layout and formatting are lost in the process. For spreadsheet analysis with precise calculation results, the Code Sandbox is more reliable because it processes the file directly instead of reading it as text only.
Note
Only JPEG, PNG, WEBP, and GIF are processed directly as images by the AI models. Vector formats such as SVG aren't among them. There's also a limit on the combined size of all attached images – see Image too large for details.
Note
Non-image files (PDF, Word, txt, …) are limited to 50 MB per file. A larger file is rejected right at upload with the notice "The file is too large. Maximum size is 50 MB", before it's ever sent to the model.
Add data source – grounds the answer in a source set up in your workspace, such as a knowledge base or a connected drive. This entry appears once data sources exist in your workspace; they are set up by your administration – more on this under Data Administration.
Prompt library – inserts a saved prompt from the Prompt Library straight into the input area. If the prompt has variables, the library asks you for the values and substitutes the placeholders automatically. Useful for recurring requests where the structure stays the same.
Project for this chat – attaches the chat to a project so the agent can access its tables. This entry appears once your workspace has projects with tables.
Web search – allows the model to research on the internet. It formulates its own search queries, loads the relevant pages and shows the sources used with a preview (title, description, favicon) so you can immediately tell what the answer is based on. More details (number of sources, passing a URL directly, prompting tips): Web Search.
Meetings – searches your meeting recordings and transcripts so the model can draw on them when answering.
Slides – turns your prompt into a presentation, optionally based on a template. More on this under Slides.
Image generation – lets the model create images for your prompt. More on this under Image Generator.
Tables – signals to the model that it may understand, create and work with spreadsheets when you ask for it in the chat. The actual table work runs through the Code Sandbox.
Artifact mode – treats longer texts, reports or email drafts as their own document instead of a message in the history. Three settings: Auto, Enabled and Off. Active artifacts appear as a separate panel next to the chat and can be downloaded as PDF, Word, Excel, or Markdown. More details: Artifacts.
At the bottom of the menu you connect and manage Connectors – such as Google Workspace, Outlook or other connected systems – directly from the chat.
Note
Which tools appear depends on the workspace configuration. As soon as you enable a tool, it shows up as a small chip next to the input field; use the × on the chip to switch it off again.
Click the microphone icon to dictate your prompt instead of typing it. The language is detected automatically, transcribed, and inserted into the input field. Handy for longer prompts, when multitasking, or when typing isn't practical.
As with speech recognition in general, transcription can occasionally insert a word or short phrase you didn't actually say - usually during silence or very quiet audio at the end of a recording (for example, an added "thank you"). Review the transcribed text before sending, especially after longer recordings.
Activate the share toggle and generate a link to send to colleagues or to support. Recipients see the chat exactly as you shared it. Useful for best-practice libraries within a team or for support tickets where the agent needs to see the full conversation.
Input draft persistence
A prompt you haven't sent yet stays in the input field even after a page reload. The draft is stored locally in the browser. Particularly useful for long prompts or when F5 was hit by accident or the tab was closed.
Sources with preview
When the model uses web search, the sources appear with title, description and favicon. You can tell at a glance whether a source is trustworthy, without having to open it.
Rotating placeholder suggestions
The empty chat input shows rotating placeholder suggestions like "Summarize this PDF" or "Build me an Excel breakdown". They lower the entry barrier for new users and give inspiration when you don't know where to start.
Context window indicator
On the right of the input field, a small circular indicator shows how much of the currently selected model's context window the ongoing chat has already used. Clicking it opens the details: tokens used vs. available, plus the approximate character count. The color shifts as usage grows (orange from 50%, red from 80%), so you can switch models or start a new chat before running into the "Chat too large for the model" error.
