Improve your bot's answers

Find the problem you're seeing below, then work through the causes in order: the first ones are the most common and the quickest to check. Each fix links to the page that owns the setting.

Who can do this: team owners and admins can change the settings below.

For why these problems happen (how knowledge, instructions, history and the model combine into a reply), read How your bot answers first. To catch problems before customers report them, see After launch.

Retrieval settings live in General settings under Model settings (advanced). Each is explained in Retrieval settings and the settings reference.

The bot says "I don't know" when the answer is in your content

1. The content hasn't synced. Check the Data sources tab: the source must have finished syncing. If a sync failed, see Troubleshooting.
2. The content isn't really there. Search your data source for the answer. Website sources can miss content that's loaded by JavaScript, or capture only part of the page if the content selector is too narrow. See Add a website.
3. Retrieval is too strict. Lower Document relevance so more content counts as a match, and raise Max context documents so the bot sees more pieces at once. Short entries, such as one-line FAQs or product rows, often need a lower relevance.
4. Follow-up questions lose the thread. Turn on Enhanced knowledge retrieval. It rewrites a follow-up like "how much is it?" into a complete question before searching. It uses more tokens.
5. The Prompt is too restrictive. Rules like "only answer questions about billing" can make the bot refuse things it knows. See Write your bot's instructions.

The bot gives wrong or made-up answers

1. Your content is out of date or contradicts itself. Remove old pages, and turn on automatic syncing. See Keep your bot up to date.
2. The Prompt allows guessing. The default prompt lets the bot fall back on the model's general knowledge when your content doesn't cover a question. To stop that, add: "Only answer from the provided context. If the answer isn't in the context, say you don't know and offer to connect the customer with the team."
3. Retrieval is too loose. Raise Document relevance so loosely related content isn't passed to the bot.
4. Creativity is too high. Lower Creativity (temperature) for more literal, consistent answers.
5. The content is hard to match. Clear headings, one topic per page and question-style titles make the right content easier to find. See Structure your content for good answers.
6. The model isn't up to the task. On plans with model selection, try a more capable model. See Models.

The bot answers questions it shouldn't

1. The Prompt doesn't set boundaries. Say what the bot is for and what to do with anything else, for example: "You only help with questions about Acme's products. For anything else, politely say you can't help with that."
2. Your knowledge covers more than you want. Remove data sources or pages the bot shouldn't talk about.
3. Creativity is high. Lower Creativity (temperature).

Answers are too long

1. The Prompt doesn't set a length. Add: "Keep answers to two or three sentences unless the customer asks for more detail. Use bullet points for steps."
2. The bot is given too much content. Lower Max context documents so it sees fewer pieces at once.
3. The model is wordy. Some models write longer answers than others. See Models.

The bot replies in the wrong language

1. The Prompt doesn't say which language to use. Add: "Always reply in the same language the customer writes in." If your Prompt is written in one language, the bot may default to it.
2. The widget language is a separate setting. The web channel's Language setting changes the widget's buttons and labels, not the bot's replies. See Languages.
3. Your content is in another language. The bot can answer in a different language from your content, but answers are most reliable when the content and the question share a language.

The bot forgets what was said earlier

1. The conversation is longer than the bot's history. The bot sees up to Max message history earlier messages (default 12, up to 50). Raise it in Model settings (advanced). More history uses more message tokens per reply.
2. The context budget is full. If Max context size (base tokens) is low, older messages are dropped to make room for knowledge. Raise it, or lower Max context documents.
3. Details should come from your site, not the chat. If the bot keeps asking for things you already know, such as the customer's name or plan, pass them in with the JavaScript SDK's identifyUser or a pre-chat form. The bot sees them on every reply.

The bot doesn't use a power-up when it should

See My bot never uses the power-up. The fix is usually a clearer Description on the power-up, a line in the Prompt saying when to use it, or a model that supports power-ups.

Check the fix worked

Click Test on your bot's page and ask the original question again, plus a couple of rewordings. On the Enterprise plan, save the question as a test case so you're told if the problem comes back. See Test your bot's answers.