Skip to main content
An assistant is what answers questions. Each assistant has its own knowledge, model and guidance, so one organisation can run several: a public documentation assistant and an internal engineering assistant, for example, with different content and a different tone. An assistant answers only from the collections attached to it, so the knowledge has to exist before the assistant is useful: how your content becomes answerable covers collections, sources and ingestion end to end.

​
Settings

Changes apply to the next question. There is nothing to restart or republish.

​
Create an assistant

  1. Open Customizations and select New assistant.
  2. Enter a Name, and optionally a Description, Model and Guidance.
  3. Under Knowledge, tick the collections the assistant should answer from. When your organisation has exactly one collection, it is already ticked. A collection embedded with a different model from the ones you have ticked cannot be combined with them and is shown disabled, with the reason.
  4. Select Create assistant.
To change an existing assistant’s knowledge, select Knowledge on its row, change the ticked collections and select Save knowledge. The row shows No knowledge attached while an assistant has none.If your organisation has no collections yet, the dialog says so. Connect a source first; the console creates a collection for it, which you can then attach.
An assistant with no knowledge attached cannot answer. Questions to it are refused with 400 validation_error until at least one collection is attached.

​
Choose a model

List the model catalogue to see which models your deployment can use, then set model to one of the returned id values. The catalogue lists only models that can answer questions, from providers your deployment is configured for. Saving an assistant with a model known not to answer, such as an embedding model, is refused with 400. The model changes how answers are written. It does not change what they are grounded in: every model answers from the same retrieved evidence.

​
Write guidance

Guidance shapes tone, format, audience and framing. Good guidance is short and specific:
Guidance cannot loosen grounding. An instruction such as “answer from general knowledge when the docs don’t cover it” has no effect: the platform checks every answer against the retrieved evidence, whatever the guidance says. See How answers work.

​
Lifecycle

Assistants cannot be deleted. Archive an assistant you no longer need, and set status back to active to bring it back. While an assistant is disabled or archived, its conversations stay readable: you can list and read them, list their messages, rename, archive or delete them. Asking a new question in one returns 400. A widget embed that uses the assistant is stricter: a visitor’s requests about a thread, reading it included, return 400. Set the assistant back to active to ask again and to bring the widget back.

​
Who can configure assistants

Owners, Administrators and Editors can create and change assistants. Members can see assistants and ask questions. See Organisations and roles.
Last modified on November 2, 2026