Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Advanced settings

The collapsed Advanced settings panel at the bottom of an assistant's settings screen. These control access, conversation limits, answer review and what happens when something goes wrong.

Open it with the Advanced settings button. You will not need most of these for a straightforward assistant.

Access

Only visible to

Restricts the assistant to one or more user groups. Leave it empty and everyone in your organisation who can see assistants can use this one.

This is the main tool for keeping an assistant contained. An assistant built on HR documents should normally be restricted to HR, not open to the whole organisation.

Group restrictions interact with sub-agents. If a restricted assistant is used as a sub-agent, users outside the allowed groups cannot reach it. The supervising assistant answers the question itself instead of handing over. The sub-agent row shows a Restricted access badge when this applies.

Conversation behaviour

Setting
What it does

Interaction interval for updates

How many exchanges pass before the assistant refreshes the conversation's topic and summary. Minimum 1.

Session size

The maximum number of messages kept in one conversation.

Assistant context size

How many of the user's earlier messages the assistant can still see. Lower values keep it focused and cheaper; higher values let it remember more of a long conversation.

Deadline for resuming a session

How long, in seconds, a visitor can come back and continue where they left off. 86400 is one day; 0 disables resuming. Applies only to an assistant embedded on a website.

Prompt template

Controls how much scaffolding the platform puts around your Instructions for AI.

Option
What it does
When to use it

Base assistant

Recommended. Builds a complete assistant around your instructions: tells the model it is helping a person, applies the assistant name and your chosen tone of voice, asks for readable formatting, and adapts length and detail to how knowledgeable the reader seems.

Almost always.

Bare prompt

Minimal framing. The model gets only the assistant name and your instructions: no role, no tone of voice, no length guidance, no formatting rules.

When you want to shape all of the behaviour yourself.

Instructions as full prompt

Your instructions become the entire brief, exactly as written, with nothing added at all.

Advanced. Only when you want complete control over the whole prompt.

Start with Base assistant. Move down the list only if it is getting in the way of something specific you are trying to do.

Handover notifications

Only relevant when Chat with an employee (Direct Chat add-on) is on.

Setting
What it does

Direct Chat email notification

E-mail addresses that get a notification when a user asks to speak to a person.

Enable email notifications (requires Direct Chat)

Switches those notifications on.

Privacy

Setting
What it does

Include user metadata in chat context

Lets the assistant see information about the signed-in user, so it can personalise answers.

Block message if pseudonymisation fails

Stops a message being processed if personal data could not be masked. Only available when Mask personal data is on in the main settings.

Memory

Allow users to save memory across conversations (beta) lets each user build up a small set of facts the assistant remembers between conversations: how they like answers formatted, which department they are in, and so on.

Memory is per user: one person's memories are never visible to another. Users manage their own list under Preferences in the chat, where they can add, edit and delete entries. There is a limit on how many entries a person can keep, and the interface warns them as they approach it.

Memory needs a capable model. If the selected model does not support it, the settings screen says so.

Answer review

You can have a second model check the assistant's answers before they reach the user.

Setting
What it does

Judge prompt

Describes when an answer should be accepted, when it should be expanded, and when it should be rejected and replaced.

Judge model

Which model performs the check. Only appears once you have written a judge prompt.

Write the judge prompt as acceptance criteria: "Accept the answer if it cites at least one policy document and does not speculate about individual contracts. Reject it if it gives legal advice."

This adds latency and cost to every answer. Use it where a wrong answer is expensive, not everywhere.

Fallbacks

Setting
What it does

Don't fail if primary model is down

Lets the assistant fall back to another model instead of failing when its usual one is unavailable.

Fallback strategy

Which fallback configuration to use. Appears once the option above is on. Global default uses your organisation's standard.

Fallback configurations are managed centrally. See Fallbacks.

Some of these settings only appear if you hold the corresponding rights. If a setting described here is missing from your screen, you do not have rights to it.

Last updated

Was this helpful?