Chatbots
Learn how to create, configure, and optimize your AI chatbots.
Creating a chatbot
Navigate to Chatbots → New Chatbot from your dashboard. Each chatbot is independent — it has its own data sources, configuration, and conversation history. Most teams create separate chatbots for different use cases (e.g., "Sales Assistant" and "Support Bot").
Presets
Presets control how the AI behaves when generating responses:
Balanced
The default preset. Provides helpful, conversational responses while staying grounded in your data sources. Best for general customer support.
Accurate
Sticks closely to information in your data sources. When the bot can't find a relevant answer, it will say so rather than improvise. Best for technical or compliance-sensitive contexts.
Detailed
Gives longer, more thorough answers with additional context and explanations. Best for technical support, onboarding guides, and documentation-heavy use cases.
System prompt
The system prompt is the high-level instruction that frames every response. It runs before retrieval, so use it for tone, persona, scope, and refusal rules — not for product facts (those belong in data sources). A good system prompt usually answers four questions:
- Who is the bot? e.g. "You are Atlas, the support assistant for Acme Tools."
- What can it do? What scope of questions is in-bounds — billing, product, returns?
- What can't it do? e.g. "Never quote prices that aren't in the data sources. Never invent SKUs."
- When to hand off? e.g. "If the user asks for a refund or escalation, suggest they reply with 'agent' to talk to a human."
Edit the system prompt under Chatbot → Customize → Behavior. Keep it short (under ~300 words) — long prompts crowd out retrieved context.
Model selection
Every chatbot generates responses with OpenAI. Pick the model under Chatbot → Customize → Model:
Faster models
Lower first-token latency and good general support quality. The default for new chatbots.
Stronger models
Better on long, nuanced answers and instruction-following. Useful when answers need to stay tightly within a long policy document.
The model you pick doesn't affect billing — every response charges the same credit cost based on the chatbot's preset (see Plans, limits & credits).
Retrieval tuning (advanced)
Each chatbot has a retrieval pipeline that searches your data sources, ranks results, and packs the best matches into the AI's context. The defaults work well for most cases, but Chatbot → Customize → Retrieval lets you tune the knobs when needed:
| Setting | Default | What it does |
|---|---|---|
| topK | 5 | How many chunks to send to the model. Higher = more context, slower, and pricier. |
| overFetchK | 15 | How many candidates to retrieve before reranking. Higher = better recall, slower. |
| minScore | 0.35 | Drops candidates below this hybrid-search score. Raise to be stricter. |
| rerank | on | LLM reranks candidates by relevance before packing. Disable to cut latency. |
| qaBoost | 1.2 | Multiplier applied to Q&A data sources. Raise to favor curated answers. |
| contextTokenBudget | 3000 | Max tokens of retrieved content sent to the model. |
| rewriteFollowUps | on | Rewrites short follow-ups ("how much?") into standalone queries before searching. |
When to tune. If the bot misses obvious answers, lower minScore or raise topK. If it hallucinates, raise minScore and switch the preset to Accurate. If responses feel slow, turn off rerank and lower overFetchK.
Welcome message
The welcome message is shown to visitors when they open the chat widget. Use it to set expectations — e.g., "Hi! I'm an AI assistant for Acme Corp. I can help with product questions, billing, and technical support. How can I help?" A good welcome message reduces friction and guides users to ask the right questions.
Chatbot analytics
Each chatbot has an Analytics tab showing message volume over time, response times, and conversation trends. Use this data to identify gaps in your knowledge base — if the bot frequently can't answer a topic, add a data source for it.
Next steps
Train your chatbot with your own content in the Data Sources guide.