Fine tuning and running costs

The Fine tuning card at the bottom of Answer Desk > AI provider already has sensible values. Change them only when something specific is wrong, then click Save changes in that card.

The settings

Setting	Default	What it does
How freely it writes	Low (30 of 100)	Low keeps answers close to your documents and consistent. Keep it low for support.
Longest answer	700	Maximum length in tokens (a token is roughly three quarters of a word), from 100 to 4000. It also caps the cost of one answer.
Extracts sent with each question	5	How many pieces of your documents go to the AI with each question, from 1 to 20. Raise it if the assistant says it does not know something that is written down. More extracts cost more.
How close an extract has to be	20 of 100	Extracts less similar than this are dropped. Too high and the assistant starts saying it does not know.

Choosing a model

In the Who answers card, Model is set to The recommended one. A bigger model writes better and costs more per question. For a support desk answering from your own documents, the smaller models are usually enough, because finding the right paragraph is the index's job.

What it costs

• You pay only your AI provider. As a rough guide, a support question on a small model costs a fraction of a cent, and indexing a whole site is usually a few cents once.
• Start here shows answers today, so you can see usage at a glance.
• Each test in the Try it box costs the same as a real question.

Put a ceiling on spending

On Answer Desk > Rules and privacy, in What is kept, and for how long:

• Messages per visitor per hour: default 20. Stops one person from spending your budget. 0 removes the limit, which is not a good idea on a public site.
• Most answers in a day: default 0 (no ceiling). A limit for the whole site. Setting one for the first month is a good idea. When it is reached, visitors are told the assistant will be available again tomorrow.