How an answer is made

Knowing the steps makes every setting, and every weak answer, easier to understand.

1. The question is made to stand on its own#

A follow-up such as "and what does that cost?" means nothing by itself. Before searching, the question is rewritten using the last few messages so that it carries its own subject. Only the search uses the rewritten version. The visitor's own words are what is stored and what the model answers.

2. Your content is searched#

Every source was cut into passages of about 500 tokens when it was added. The question is compared with all of them by meaning, not by keyword, so "how long may I send it back?" finds your returns policy even if the word "return" is not in the question.

When the visitor asks from a page of your site that the agent has read as a source, the 2 passages of that page closest to the question are added first, whatever their score: "is this suitable outdoors?" on a product's page is about that product, however little the words match. Only the launcher script can tell the agent the page, see Install the widget.

3. Confidence is measured#

Before anything is written, the search result itself is scored from 0 to 100. The score looks at how good the best passage is and at how far it stands out from the rest: a question your content covers has one clear winner, a question it does not cover finds ten passages that are all equally unlike it.

• 70 or more: well covered.
• Below 50: your content had little to offer. The fix is a source.

Confidence is always measured the same way, whatever your retrieval settings are, so a number means the same thing on every agent and over time. The passages added for the visitor's page do not count toward it.

4. The model writes#

The model receives the platform's rules, your instructions, the page the visitor is on when it is known, and the best passages, up to the number you set. It is told to answer only from those passages and from facts in your instructions, to say so when they do not contain the answer, and to stay on your subject.

5. A second opinion#

For answers the search was confident about, a separate check follows. A model that has seen neither your instructions nor the conversation reads the answer next to the passages and gives a verdict:

• supported: every factual claim is backed by the passages
• partial: some claims are, at least one is not
• unsupported: the main claim is not in the passages, or contradicts them

It is only asked about confident answers, because that is where an unsupported one would otherwise hide behind a good number. Below 50 no check is made and grounding is empty. The check happens after the visitor has their answer, so it never slows a conversation down.

6. It is recorded#

The answer, its confidence, its verdict, the sources it used and what it cost are stored with the conversation. That is what Activity and Overview show you.

Reading the two together#