Assistants
An assistant is a chat experience that answers from a dataset. When someone asks a question, it searches the connected dataset for relevant content, grounds its answer in what it finds, and can show the sources it used — so answers stay tied to your knowledge rather than the open internet.
What an Assistant Needs
- A name.
- A connected memory dataset — its memory.
- Optional behavior instructions — the tone and the rules it follows.
- Optional security settings for embedding or sharing.
What You Can Shape
Behavior, search settings, appearance, hints, translations, and forms — plus the keys and allowed origins used when you embed it. Everything is tested in the playground before it goes live, so nobody sees a half-configured assistant.
For example, an online store gives its assistant one rule — "Answer only from our catalog and policies; if you're unsure, tell the customer to contact support" — connects it to the product dataset, and embeds it as a floating button. A visitor asks "do you ship to Canada?"; the assistant answers straight from the shipping policy page and links to it. Change the store's shipping terms in the dataset, and the next answer reflects it — no edit to the assistant required.
An assistant is only as good as its dataset
If answers look weak, the fix is almost always in the dataset — missing content, or a page the data capture job didn't reach — not in the assistant itself.
Go Deeper
- How Configuration Is Organized
- Configure Assistant Behavior
- Configure Assistant Search
- Security
- Embed an Assistant on a Website