Browse the docs

Instructions, models, and how it thinks

What to put in the instructions, how to pick a model, and the settings under Configure that change cost and quality.

On this page

The Configure tab is where an assistant's behavior is set. The instructions, the model, and the sources decide most of the result.

Instructions#

Write them the way you would brief someone on their first day. Who is it talking to? What must it always do? What must it never do? Short and specific works better than long and general.

Configure
NameCustomer support guide
InstructionsAnswer only from the sources. If they do not cover the question, say so and suggest emailing support. Never promise a delivery date or a price that is not in the sources. Keep answers under 120 words unless asked for detail.
ModelClaude Haiku 5.5
Save

The three sentences in that sketch work on almost every assistant. Every change to the instructions is saved as a version, so you can go back.

Choosing a model#

Models differ in quality, speed, and price, and the ranking changes every few months. Rather than guess, use Compare models. It asks your own checks of several models and tells you which gets them right for the least money.

For most support and sales assistants a fast, cheap model does as well as an expensive one. The answers come from your sources, not from the model's memory.

You can also set a fallback model under Settings → AI columns. It answers when the main one is overloaded, rate limited, or too slow.

Settings that change cost and quality#

They sit under How it finds answers, thinking, and cost on the Configure tab. The defaults suit most assistants.

  • How it uses your data. Automatic is the one we recommend. It sends everything while it fits in one message, then switches to search. Send everything suits small data. Search by keywords sends only the parts that share words with the question. Search by meaning and keywords also finds parts worded differently (it needs an OpenAI or OpenRouter key). Agent searches, opens, and queries your sources step by step and shows each step. It is best for large folders, at several model calls per answer.
  • Thinking before answering. Quick, Balanced, or Thorough. It helps with open questions, but it is slower and the thinking is billed.
  • Let the model write its own searches. It turns the question into its own search words first. Better on vague questions, at one extra model call per answer.
  • Data each message can use. The most of your data sent with each question. When it all fits, the assistant reads everything.
  • Skip repeated text. Text found in several files is sent once.
  • Reuse what was just sent (prompt caching). It cuts the cost of follow-ups by up to 90%. Leave it on.
  • Passage size. Smaller passages are more precise and cheaper. Larger ones keep more context. Changing it re-splits all your data.

Seeing what it cost#

The Insights tab, next to Chat, shows questions by day, what they cost, which ones it could not answer, and which were asked most. The questions it could not answer are your list of sources to add next.

Updated Nov 3, 2026