---
title: "The room assistant"
description: "The assistant visitors chat with inside a room: what it answers from, how it cites, how it treats watermarked files, and how to set its persona."
canonical: "https://undocsend.com/docs/ai/room-assistant"
last_updated: "2026-11-19"
---

# The room assistant

The assistant visitors chat with inside a room: what it answers from, how it cites, how it treats watermarked files, and how to set its persona.

## Where visitors see it

The assistant is part of every room. A visitor opens it with **Ask AI** in the room header or the floating **Ask** button at the bottom right, which carries the assistant’s name. A panel slides out with your welcome message and a few **Try asking** suggestions. The visitor types into **Ask … anything about this room**.

The conversation stays with the visitor, so they can pick it up again on their next visit.

The assistant is not shown:

- To visitors who arrived through a share link for a single document or a hand-picked set of documents. The assistant works across the whole room, so it stays out of narrower links.
- When you preview the room from the dashboard, so your own testing isn’t billed or logged.

The machine view of the room (see [Visitors’ own AI tools](https://undocsend.com/docs/ai/investor-ai-tools)) has the same assistant as a console.

## What it answers from

For each question, the assistant uses:

- The text of the room’s documents that **this visitor** may open. Files limited to another audience are left out, so the assistant can’t reveal them.
- Your [Shared Knowledge](https://undocsend.com/docs/ai/knowledge): entries for every room plus any entries for this room.
- Answers your team published to the room FAQ that this visitor may see. See [Questions and answers](https://undocsend.com/docs/ai/questions-and-answers).

If the room has nothing to answer from yet, the assistant says so plainly instead of guessing. When the documents don’t cover a question, an **Ask the team** button under the reply lets the visitor send it to you.

> Scanned documents are always made searchable with built-in text recognition. To have AI re-read pages that came out unclear, turn on **Improve scanned text with AI** in the room’s **Settings**. It uses credits or your own key, at most 20 pages per document, and applies to documents uploaded or reprocessed after you turn it on.

## Citations

When the assistant uses a document, it names it and points to the place, such as “the pitch deck, slide 12”, so the visitor can check the source. When it uses a Shared Knowledge briefing that has a source link, it links to that source. It never makes up a link.

## Watermarked files

If a visitor would see a file with a watermark, the assistant can still use it to answer: it can summarise, explain and give facts and figures from it. It won’t reproduce more than a short phrase word for word. If a visitor asks for the exact wording, it tells them to open the document in the room.

As a backstop, any run of 25 or more words copied from a watermarked file is replaced in the reply with `[…see the document in the room]`. The stored transcript shows the same replacement. Text from files without a watermark is not affected.

## Set the persona

The persona is the workspace default every room inherits. To change it, go to **Agent** and click **Persona**, edit the fields, and click **Save agent settings**.

- **Name**: The assistant’s display name. Visitors see it on the button and in the panel.
- **Role description**: Completes “You are NAME, …”. For example, “an AI analyst embedded in the dealroom”.
- **Identity paragraph**: Grounds the assistant: who is behind the room and which questions are on-topic.
- **Voice rules**: Tone and style. One rule per line.
- **Hard rules**: Things it must never say or do.
- **Attribution label**: The brand it names when asked who built it. When set, the panel reads “NAME · powered by LABEL”. Attribution URL is an optional link next to it.
- **Welcome text**: The first message visitors see when they open the assistant.

Under **Advanced** you can set a daily message limit per visitor (see below), the longest message a visitor may send (**Max input chars**), the longest reply (**Max output tokens**), and how documents are fed to the assistant (**Chat mode**). The defaults suit most rooms.

The models that write and screen answers are chosen on the **Agent** console. See [AI credits and your own key](https://undocsend.com/docs/ai/credits-and-your-own-key).

## Override it for one room

A room can sound different from the rest of your workspace. Overrides apply to that room only, and only owners and admins can set them. Other team members see the room’s settings read-only.

1. Open the room and click the **Agent** tab.
2. Each field shows **Inherited** and the workspace default. Click **Override for this room**.
3. Edit the value and click **Save override**. The field now shows **Overridden**.

To go back to the workspace default, click **Revert to default**.

You can override **Voice & tone**, **Hard rules**, **Identity**, **Welcome message**, **Role**, **Answer model** and **Screening model**. The name, attribution and advanced settings always come from the workspace.

### Things to know

- A room’s **Hard rules** replace the workspace hard rules in that room. Copy across any workspace rules that should still apply.
- A room’s model choice replaces the model on your credits step. Your own-key steps keep their own models unless the room’s choice is from the same provider.
- The room’s overrides also shape the AI drafts for visitor questions asked in that room.

## Message limits

By default there is no daily limit. To set one, go to **Agent → Persona → Advanced** and enter **Daily messages per visitor**. It counts each visitor’s messages across all your rooms and resets at midnight UTC. Visitors who reach it see your **Rate-limited reply**.

## Pausing the assistant

There is no switch to hide the assistant from a room. To stop it spending, set a budget cap on the **Agent** console and choose **Pause the agent for the rest of the cycle** as what happens when a cap is hit. See [budget caps](https://undocsend.com/docs/ai/credits-and-your-own-key#budget-caps).

When the assistant can’t answer, for example because credits ran out, visitors see a short note that it isn’t available right now, that their access is unaffected, and that they can reply to whoever shared the room. It never tells them why.

---

Section: AI. HTML: https://undocsend.com/docs/ai/room-assistant · Docs index: https://undocsend.com/docs.md
