# Chat Start a conversation, pick a model, and steer how the assistant answers. Open the app and a new conversation is waiting. Type your message and press **Enter** to send it. **Shift+Enter** adds a line break. The controls sit along the bottom of the message box: the **+** menu and web search on the left, the model picker, dictation and send on the right. ## Starting a new conversation **New Chat** in the sidebar starts a new conversation. It stays a draft, with no entry in your history, until you send the first message. Then it is saved and gets a title automatically. ## Picking a model The model picker lists every model your organisation has turned on, each with a one-line description. The description is either a summary your platform admin wrote or a line built from the model's maker, size and context window, such as `Alibaba · 235B · 128K context`. There's usually a fast everyday model marked as default and heavier models for harder work. A model under maintenance is greyed out and shows the maintenance note. The app remembers your choice on each device. If that model is removed or goes into maintenance, the app switches to the default, so you're never left on a model that can't answer. ### Switching models in a conversation You can change model at any point, and the next reply comes from the new model. To answer, a model reads the whole conversation from the start. Each model keeps the conversations it read recently in a cache, and input read from that cache costs a fraction of the normal input rate. The cache belongs to one model. So the first message after a switch pays the new model's full input rate for the whole history. After that the new model reads from its own cache, and if you switch back to a model you used recently, its cache is often still there. If the conversation is longer than the new model's context window, the model doesn't read the earliest messages. Nothing is deleted: the messages stay in the conversation, and a model with a larger window sees them again. ## Reasoning (thinking) toggle Some models can reason before they answer, spending extra tokens to work through the problem first. When the selected model can either think or answer directly, a **Thinking** button appears in the message box. Click it to switch thinking on or off. - **On**: the reply starts with a collapsed thinking trace. Expand it to see the model's working. Good for code, analysis and multi-step problems. - **Off**: the model answers directly. Faster and cheaper, and fine for short questions. The app remembers your choice for each model, so a heavy model can stay in thinking mode while a fast one answers directly. ### How the thinking trace renders While the model reasons you see **Thinking** and a running timer. When it stops, the trace collapses to a summary such as **Thought for 12s** and the answer appears below it. If the model reasons again later in the same reply, after calling a tool for example, the indicator comes back for that step. ## Web search If your organisation has web search turned on, click the **globe** before you send a question that needs current information, such as news or prices. The assistant searches the web and cites what it finds in its reply. While web search is on, the globe shows as a **Web Search** pill. Click it again to turn it off. Web search is off by default and never runs unless you turn it on. Your org admin decides whether web search is available. If it reads "Not enabled for your organisation", it isn't available to you. It is also off in a conversation connected to a private source, and reads "this chat is isolated". ## Code execution If your organisation has code execution turned on, the assistant can write and run code in an isolated sandbox to finish a task: crunch a spreadsheet, draw a chart, convert a document. The sandbox is created for the conversation and cleaned up afterwards. You don't manage it. The sandbox has small, fixed resource limits, and the assistant knows them. Ask for something clearly too big, such as tens of gigabytes of files, and it tells you so up front or splits the work into pieces that fit. If a running job goes over a limit, the sandbox session is stopped and the files and variables in it are lost. So when the assistant says a job won't fit, make the job smaller. Your org admin decides whether code execution is available. The sandbox limits are fixed and can't be changed. ## Attaching files You can attach files three ways: 1. **Add files or photos** in the **+** menu. 2. Drag and drop files onto the chat. 3. Paste an image or a large block of text into the message box. You can attach PDFs, Word, PowerPoint, plain text, CSV and Excel, JSON, code files and common image formats. [Files and sources](https://docs.privatemind.com/files-and-sources.html) explains what happens to each type. Each file shows as a chip in the message box, with its name, its type and a progress bar while it uploads. A file the platform can't take turns red and reads **Failed**. Click **✕** on a chip to remove the file before you send. When you send, the chips move above your message in the conversation, so you can see which files went with which message, even after a reload. Each message shows only the files you attached to it. An attached file is also added to the conversation's sources, so the assistant can keep using it in later turns. A paste of 15,000 characters or more, or your model's inline limit if that is lower, becomes a `.txt` source and is attached. A paste of 1,500 characters or 10 lines or more shows as a chip in the message box. It is still sent to the assistant in full, and you can click the chip to view or edit it. ## Mentioning connectors with @ Type `@` in the message box to attach a connector from your library, an MCP server or a REST API, to this conversation. **↑** and **↓** move through the list, **Enter** attaches the selected connector, and **Esc** closes the list. Once a connector is attached, the assistant can use it on every turn. The **Connectors** panel lets you switch connectors off and on without removing them. ## Voice input Click the **microphone** to dictate. Your speech is typed into the message box so you can check it before sending. When the message box is empty, the send button becomes a **voice mode** button. Voice mode is a hands-free conversation in which the assistant speaks its replies. Press **Esc** to leave it. Both use your browser's microphone permission. If nothing happens when you click the mic, check the site's permission in your browser. ## Ephemeral chats Turn on **Ephemeral mode** in the sidebar, under New Chat, and every conversation you start is ephemeral: - Its messages, attachments, vector collections and sandboxes are **permanently deleted** when your session ends: when you sign out, your token expires or the session is lost. - Connectors you added from your library are unlinked from the conversation but stay in your library. - A violet ring around your avatar and a banner in the header show that ephemeral mode is on. Use it for sensitive material you don't want kept. It's the chat's closest thing to private browsing. ## Previewing what the assistant builds When the assistant makes something self-contained, such as a small web page or app, a chart, a diagram or a document, it appears in the reply as an **artifact card**. Click **View** on the card to open it in the artifact panel. Most files preview in the panel without a download: - **Web pages and apps**: HTML runs live in a sandbox, scripts included. The card reads **Website**. - **Charts and graphics**: SVG and image files (PNG, JPG, GIF, WEBP and others) show as the picture. - **Documents and notes**: Markdown renders formatted. PDFs open in the built-in viewer. - **Office files**: Word (`.docx`) and Excel (`.xlsx`) files keep their formatting and tables. PowerPoint (`.pptx`) files show the slide text only, without the layout or graphics. - **Data**: CSV files show as a table. - **Code and text**: shown with syntax highlighting and a copy button. Previews work in older conversations too: reopen the chat and click **View**. A file the panel can't show gets a **Download** card. For an exact copy of any Office file, download it and open it in the desktop app. Previewed web pages run isolated from the app, so a generated page can't reach your session or your data. You can click around an interactive page the assistant built safely. ## Stopping a response While the assistant is writing, the send button becomes **Stop**. Click it to end the reply. What it wrote so far is saved: you can edit your prompt and try again, or carry on from there. ## Regenerating a response To get a fresh answer, hover over the latest assistant reply and click **↻ Regenerate**. Choose the same model or a different one. The new answer replaces the old one in place, and the old answer is kept as a version you can switch back to (see [Switching between answers](https://docs.privatemind.com/chat.html#switching-between-answers)). Only the latest reply can be regenerated. While it streams, the send button becomes **Stop**. ## Editing a message To change something you sent, hover over your message, click the **✏️ pencil**, edit the text and choose **Save & Regenerate**. The assistant answers again from that message, using the model now selected in the picker, and the replies that came after it are replaced. Your original message and its replies are kept as a separate branch you can go back to. ## Switching between answers Regenerating or editing keeps the previous version. A message with more than one version shows a counter such as **2 / 3** with arrows. Use the arrows to step through the versions. This works on your edited messages and on the assistant's regenerated replies. ## Where next - [Files and sources](https://docs.privatemind.com/files-and-sources.html): what happens to documents you attach - [Conversation history](https://docs.privatemind.com/conversation-history.html): rename, share and export past chats - [Projects](https://docs.privatemind.com/projects.html): pin a set of sources to every new conversation - [Automations](https://docs.privatemind.com/cloud-agents.html): run the assistant and other steps on a schedule, from a webhook or on demand