Skip to content
English
  • There are no suggestions because the search field is empty.

Fine-Tuning AI Behavior with Chat Settings

Advanced controls for how your bot thinks, searches, and responds. Most bots work well right out of the box. Use Chat Settings when you want more control over response length, how strictly the bot matches your knowledge base, or the exact instructions it follows.

Access Chat Settings

Navigate to Chat Settings in the sidebar. Settings are specific to each bot.

Response quality

- AI Model: choose between model tiers optimized for different response complexity and latency.
- Timezone: sets accurate timestamps for that bot's conversations.
- Max Tokens (64–4096): controls how long responses can be — higher allows more detailed answers.

Knowledge base retrieval

- Top K Chunks (1–10): how many knowledge sources the bot considers per answer. Higher considers more sources, but answers may be less focused.
- Top K FAQ Chunks (1–10): how many FAQ suggestions the bot considers.
- Similarity Threshold (0.1–1.0): how closely a source has to match the question. Higher means stricter matching, and a lower chance the bot answers from something only loosely related.

Conversation memory

- Conversation History Window (0–20): how many previous messages the bot factors in as context. Higher gives more continuity in longer conversations, at the cost of more processing per response.

AI Prompts (advanced)

For full control, edit the instructions the bot follows for each task:

- Classifier Prompt — how it categorizes incoming questions
- Conversational Prompt — general conversation behavior
- Topical Prompt  — topic-specific answer behavior
- Conversation Trends Prompt — how it analyzes common question topics
- FAQ Suggestions Prompt — how it generates FAQ recommendations

Click Save Settings when you're done — changes persist across refreshes and apply immediately to that bot.

💡 **Pro tip:** If you want the bot to stick closely to your documented content rather than answering from general knowledge, raise the Similarity Threshold. It won't fully prevent general responses when nothing in your knowledge base matches, but it makes near-misses far less likely.


---
Related: Building Your Bot's Knowledge Base · Reviewing Conversations & Bot Performance