Answer quality
Workspace → Agent settings. Every setting here is per agent, so two sites in the same workspace can sound nothing alike.
Persona and tone
Section titled “Persona and tone”A persona preset sets the assistant’s default posture: sales, support, consultant, concierge, or the neutral default. It changes how it opens, how much it pushes, and what it does when a question is outside its knowledge.
Formality decides between the formal and informal second person in
languages that distinguish them — critical in German, and set to auto by
default so it follows the visitor.
Extra instructions are appended to the system prompt verbatim. This is where house rules go:
Never quote delivery times. If asked, say it depends on the destination and offer to connect them with sales.
Keep them short and absolute. Instructions that argue with each other produce answers that pick one at random.
Strictness
Section titled “Strictness”Strictness governs what happens when the retrieved pages do not really cover the question.
- Strict — say so and stop. Best where a wrong answer is expensive: regulated services, prices, eligibility.
- Balanced — answer from what was retrieved, and flag what is uncertain.
- Loose — general knowledge is allowed to fill gaps around your content. Fine for a blog, dangerous for a price list.
Retrieval depth
Section titled “Retrieval depth”Top-K is how many passages are pulled in and offered to the model.
More is not better. A larger K brings in loosely-related passages that compete with the right ones, and long, hedged answers are the usual symptom of too much context rather than too little. Raise it when answers are missing facts that are demonstrably on the site; lower it when answers wander.
Languages
Section titled “Languages”Localization lets the assistant reply in the visitor’s language, whatever your site is written in, and translates the chat interface around it too. The translations are cached, so each new language costs one call the first time anyone uses it and nothing after.
The default language is what an undetectable visitor gets. Your authored texts — greeting, teasers, suggested questions — are written in that language and translated on demand.
The model
Section titled “The model”Provider is auto unless you have a reason. Auto uses the best configured
provider and falls back when one is unavailable, which is what you want on a
Tuesday afternoon when someone else’s API is having a bad day. Pinning a
specific model is for reproducibility, not quality.
Effort is chosen per message: a short factual question with strong retrieval gets a fast answer, an ambiguous one with weak retrieval gets more thinking. You do not configure this.
The greeting and the questions
Section titled “The greeting and the questions”- Initial message — the first thing in the panel when it opens.
- Suggested questions — clickable starters. Make them the questions you actually want asked, not the ones you wish were easy.
- Teasers — a bubble that appears next to the launcher after a delay. It raises engagement and it irritates people; the delay is the dial between the two. Mobile can differ from desktop: keep the full set, show a shorter custom set, or switch teasers off on phones entirely (Deploy → Chat widget → Teaser bubbles → On mobile devices).
Checking your work
Section titled “Checking your work”Every stored message keeps the retrieval score that produced it. On a conversation, a low score under a confident-sounding answer is the thing to look at: the model wrote well from bad evidence. That is a content problem or a strictness problem, never a prompt problem.
Correcting an answer
Section titled “Correcting an answer”You correct the assistant by giving it the answer it should have given. Nothing about the model changes; what changes is what it can find.
- Open the conversation under Inbox → Conversations and find the answer.
- Below it, choose Correct this answer, pick what was wrong (not correct, incomplete, outdated, wrong tone) and say so in a sentence.
- It joins Inbox → Improvements, beside the questions the analysis found on its own. Open it there and use Add correction to knowledge. What you write is saved as a snippet titled with the visitor’s own question, and the assistant answers from it from then on.
Two things are worth knowing before you write one.
A correction is knowledge, not a rule. It competes with your pages in retrieval like any other passage, so write it as a complete answer to the question rather than as an instruction to the model. Instructions belong in Extra instructions above, and they apply to every answer.
And if the answer was wrong because your website is wrong or silent, fix the page instead and let the crawler pick it up. A snippet only helps people who reach the chat; the page helps them and everyone who arrives from search. See content gaps for the same choice made at scale.