> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.rydhm.ai/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.rydhm.ai/_mcp/server. # Chat Every expert can chat by text. Your team chats with experts inside the Rydhm app, and visitors chat on the expert's [public page](/channels/public-links) and [website widget](/channels/web-widget). Replies draw on the expert's persona and [knowledge](/knowledge/overview), and show the sources they used. ## Chat in the app #### Open a chat Go to **Experts** and select **Chat** on an expert's card, or **Chat** in its menu. #### Ask something Type into **Message (name)…** and press Enter, or select a suggestion such as **What can you help me with?** You can also dictate. #### Read the reply While the expert works you see **Thinking**, or **Reading N sources** when it searches its knowledge. Select **Stop replying** to stop a reply. From the chat toolbar you can start a **New chat**, switch to a voice call with **Talk**, send the expert to a meeting, or open **Past chats**. In-app chats can use your workspace's [tools](/experts/tools), such as checking a calendar or drafting an email, depending on what the expert is allowed to do. Some actions ask you to approve them first. Other places to chat: * **The expert editor's Test panel**: try your changes before you save. * **Ask (expert) on a meeting**: questions about that meeting, grounded in its transcript and the expert's knowledge. * **Continue in chat** on one of your own conversations in [Conversations](/quality/conversations). You can't attach files to a chat. To give an expert documents, add them to its [knowledge](/knowledge/library). ## Public chat On the public page and in the widget, visitors type into **Type a message…** and press Enter. The expert greets them first with the widget's greeting, or its first message. * **No account needed.** Visitors are anonymous. * **Knowledge only.** The expert answers from its knowledge, and can search the web if you turned that on. Your workspace's other tools never run for visitors. * **One conversation per visit.** Reloading the page starts a new chat. * **Up to 4,000 characters** per message. * **Language.** If the expert speaks several languages, visitors can switch with the [language picker](/channels/public-links#the-language-picker). Visitor chats appear in [Conversations](/quality/conversations) under **Public**, read-only. ## Sources When a reply uses the expert's knowledge, a **Sources** row under it lists up to five chips: | Source | Chip shows | Selecting it | | ---------------- | ------------------------------ | --------------------------- | | A web page | The page title, or its address | Opens the page in a new tab | | An FAQ | The question | | | An uploaded file | The file name | | Visitors only ever see web pages and FAQs as sources. The names of your uploaded files are never shown to them. Sources are saved with the message, so they're still there when you reopen a conversation. ## Safety screening Rydhm screens what **visitors** send to public experts, on share links and the website widget, before the expert sees it. It looks for: * attempts to change the expert's instructions or make it reveal them (prompt injection), in many languages; * abuse, slurs and threats; * spam, such as the same message over and over or a flood of links. Ordinary questions are never blocked for their topic: pricing, complaints and hard questions all get through. Messages that look suspicious get a closer look by a classifier model. The result is one of three: | Result | What happens | | ----------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Allowed** | The expert answers as usual | | **Warned** | The expert answers, with a hidden note to stay within its instructions | | **Blocked** | The expert doesn't see the message. The visitor gets: "I can only help with questions about what this expert covers. Ask me something else and I'll pick it up from there." | After 3 blocked messages in 10 minutes, the visitor must wait 15 minutes: "Let's take a short break. Please try again in a few minutes." Values passed into the expert's `{{variables}}` through a link or the widget are screened too. What you see: * Blocked messages and the refusal stay in the conversation in **Conversations**. * **Analytics → Visitor feedback** shows **Blocked messages** for the range. * The **Blocked visitor messages** [alert](/workspace/alerts) tells admins when blocks start. > **Note** > > Chats in the app aren't screened. Public voice calls and phone calls are checked after each reply and counted, but not blocked. If screening can't run for any reason, messages are allowed through rather than blocked. For the instructions themselves, use the expert's [guardrails](/experts/guardrails). ## Limits and cost * Chat is charged by tokens: the question, the knowledge the expert read, and its reply. A typical reply is about \$0.007. See [Billing](/workspace/billing). * The expert reads the last 60 messages of a chat when it replies. * Visitor chats are rate limited. See [Public links](/channels/public-links#limits). ## Related * [Public links](/channels/public-links) * [Website widget](/channels/web-widget) * [Knowledge overview](/knowledge/overview)