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

AI assistants

Everything an administrator needs to build, configure and maintain assistants.

An assistant is a chat-based helper with a fixed job. You give it a name, a short description, written instructions describing how it should behave, and access to the knowledge and tools it needs. From then on, anyone you give access to can hold a conversation with it.

A well-built assistant is narrow. "Answers questions about our HR policies from the staff handbook" produces far better results than "knows everything about the company". If you need to cover several areas, build several assistants and let one delegate to the others. See Sub-agents.

The pieces of an assistant

Piece
What it does
Where you set it

Identity and instructions

Name, short description, the written instructions that shape its behaviour

Messages

Opening message, tone of voice, creativity

Knowledge

The documents it can draw on when answering

Tools

Actions it can perform, like looking something up in another system

Connectors

Access to actions on external servers

Suggested questions

The starter prompts a user sees on an empty chat

Appearance

Colours, logo and shape of the chat window

Access

Which user groups may use it

Working on its own

Starts work without a chat: on a schedule, from another system or from a forwarded e-mail

Runs

What it did on its own, step by step

Assurance

Automated checks that it keeps answering correctly

The usual order of work

  1. Create it. Name, short description and opening message: three fields, thirty seconds. See Creating an assistant.

  2. Write the instructions. This is where most of the quality comes from. See Settings.

  3. Give it knowledge. Attach the collections or search tools it should draw on. See Knowledge and sources.

  4. Try it. The Preview tab shows exactly what a user will see. See Preview and embedding.

  5. Open it up. Set which user groups may use it, then turn on Assistant visible to users.

  6. Keep it honest. Add quality control scenarios so you find out when an answer drifts. See Quality control.

An assistant stays invisible to users until Assistant visible to users is switched on. You can build and test in peace before anyone sees it.

The assistant overview

AI assistants → Assistants shows every assistant you can see, as cards. Users can star an assistant so it appears at the top of the list. A disabled assistant is greyed out.

The ⋮ menu on each card is the administrator's shortcut into Settings, Tools, Quality control, Embed and Delete. Which of those you see depends on your rights.

Last updated

Was this helpful?