# ThurWiki for newcomers – the AI assistant explained > The ThurWiki AI assistant drafts, summarises, builds tables, searches the wiki, creates pages. With voice input, eight AI connections and your own rights. Quelle: https://thurwiki.ch/en/basics/ai-assistant.html Anbieter: ThurData GmbH, Siegershausen, Schweiz Sprache: Deutsch (Schweiz) --- - [ThurWiki](https://thurwiki.ch/en/basics/index.html) - AI assistant # An assistant with your rights. And no more. The ThurWiki AI assistant knows the page you have open. It drafts, summarises, translates, builds tables, searches the wiki, creates pages and finds images. And it does only what the signed-in person would be allowed to do. ## What it can do. It stands beside the page, not in another window. What it suggests, it can insert directly; what it finds, it links to. - **Drafting and shortening.** "Turn these keywords into an instruction." "Cut this section in half." "Write this more formally." - **Building tables.** Out of a list, out of an e-mail, out of an attached Excel file. Finished as a real table, sortable and filterable. - **Searching the wiki and answering.** "Where does it say how we reset the router?" It searches what you are allowed to see, and points to the page. - **Creating and filling pages.** "Create a page for Gewerbepark Nord in the Projects space, using our project template." It asks before it writes. - **Translating.** The instruction in French for the branch in Lausanne, as a new page beside the original. - **Finding and inserting images.** Through the image search across nine archives, with licence attribution, or generated when the subject exists nowhere. - **Speaking instead of typing.** A microphone button in the assistant: through the browser's own recognition or through Whisper, on your own hardware if you like. - **Transcribing voice notes.** The same speech recognition also writes out recordings, for instance a meeting you put on the page with the macro [Voice note or meeting](https://thurwiki.ch/en/basics/features.html#sprachnotiz). You see and correct the transcript before it is inserted. *Bild: The AI assistant window at the right-hand edge: the question about what to watch out for before a restart, below it the steps the assistant took – it searched and read two pages – the answer with the time and duration of the maintenance window, and at the foot the* **The assistant.** Choose a connection, tick the box, ask or speak. ## Four properties that make it useful rather than dangerous. An assistant that sees everything and may do anything is not a gain in a company wiki, it is a risk. So the following holds: ### It works with your rights. What you may not see, it does not see. What you may not change, it cannot change. HR asks it about salaries and gets an answer; an apprentice asks the same question and gets none – and no hint that such a page exists. ### Anything that writes asks first. The "may write" tick box is yours. Without it the assistant only reads. With it, it shows what it intends to do, and you decide. Every step it takes stands visibly in the conversation. ### What it does stands under your name. It is your change; the assistant only typed it. It appears in the page history like any other, with comparison and restore. There is no anonymous author called "AI" for anything to hide behind. ### A model in your own house, if that is what you want. You can set up several connections and switch between them in the assistant: a powerful model outside the house for drafting, Ollama in your own house for anything touching confidential company matters. For a model in your own house the address of your server has to be entered, otherwise the access cannot be saved. That way no question ends up at an outside service by accident. *Bild: A wiki page being edited with the AI assistant open beside it; in the conversation the instruction to turn the keywords into a table, and the answer with the finished table and the* **Beside the page.** The assistant sees what you see, and puts its result where the cursor stands. ## Eight AI connections. You choose. A connection belongs either to the wiki or to a person. The wiki administrator sets up the shared ones; whoever has a key of their own enters it in their profile and is the only one who sees it. | Claude | Through an API key; the model can be chosen | |---|---| | ChatGPT (OpenAI) | Model can be chosen; the same connection provides voice input and the transcription of voice notes through Whisper | | Google Gemini | Model can be chosen | | DeepSeek | Model can be chosen | | LiteLLM | A switch through which you can reach any model, including ones in your own house; the address of your server is required | |---|---| | Ollama | Models in your own house, without a single question leaving it; the address of your server is required | | Whisper | Speech recognition in your own house for voice input and for transcribing voice notes | On top of that comes web search (SearXNG, for instance, a search service you can run yourself), so that the assistant can also look outside the wiki. It belongs to the connections for search and images. **About voice input.** Two routes, selectable in the assistant: the browser's own recognition, which works in Chrome and Edge with nothing to set up, though as a rule the browser maker processes the recording. Or the recording that the wiki sends to Whisper, outside the house through OpenAI or on your own hardware. Anyone who does not want the recording to leave the house uses Whisper in their own house. The microphone only becomes live at the press of the button; the page never asks for it anywhere else. Voice notes too are transcribed by ChatGPT or Whisper; with Whisper in your own house that audio also stays in the house. ## Images: search, generate, attribute. The assistant and the INSERT tab share the image search: nine archives at once, every thumbnail with source, author and licence. When you insert an image, the wiki records these details with the attachment. The attribution does not get lost, and nearly all free images require it. ### Plenty even without a key Wikimedia Commons and Openverse need no key; they will get you surprisingly far. Pixabay, Unsplash, Pexels and Freepik with an account of your own. ### Licensed material visible, not insertable Hits from Adobe Stock, Getty Images and iStock are shown but cannot be inserted; the thumbnail says why. That way you find a subject and then decide whether to buy a licence. ### Thumbnails over our own server The preview images run over our own server. Otherwise every archive would learn who in the house is looking for what, before anybody has even settled on an image. ## Questions about the AI assistant **Which AI models can ThurWiki use?** Claude (through an API key), ChatGPT, Google Gemini, DeepSeek, LiteLLM (a switch through which you can reach any model) and Ollama in your own house. For voice input and for transcribing voice notes, Whisper as well, outside the house through OpenAI or on your own hardware. You can set up several connections and switch between them inside the assistant. **Does our content go to the AI provider?** The question, the page you have open and the passages the assistant looks up in the wiki for its answer – and only to the connection you have currently chosen. For confidential company matters you set up a model in your own house (Ollama or LiteLLM). For that the address of your server has to be entered; without it the access cannot be saved. That way no question can end up at an outside service by accident. **Can the assistant change or delete pages?** Change yes, delete never. It can only change what you are allowed to change, and only if you permit it: the "may write" tick box is yours; without it the assistant only reads. Every change it makes appears in the history under your name and can be brought back like any other. **How does voice input work?** A microphone button in the assistant. In Chrome and Edge the browser's own recognition can be used; there, as a rule, the browser maker processes the recording. Alternatively the wiki makes the recording, sends it to Whisper – on your own hardware if you like – and puts the text into the input field. You choose the route inside the assistant. Without a ChatGPT or Whisper connection, only the browser route is available. **What does the AI cost?** The models bill by usage, at the respective provider, through your own key. Ollama in your own house costs nothing per request but needs a capable machine of its own. [**Security and data protection** Swiss location, sign-in, intrusion detection, secrets, the log](https://thurwiki.ch/en/basics/security.html) [**All 41 connectors** bexio, WHMCS, Abacus, Jira, Git, ticket systems, Zabbix, CRM, calendar, e-mail](https://thurwiki.ch/en/basics/connectors.html) [**All the features** Editor, tables, search, versions, approvals, images, mobile](https://thurwiki.ch/en/basics/features.html) ## Ask the assistant yourself. An AI connection is set up in the demo wiki. Ask it the questions your team asks every day. [Request demo access](https://thurwiki.ch/en/basics/../demo.html) [On to security](https://thurwiki.ch/en/basics/security.html)