Improving chatbot responses

Use a strong system prompt

The system prompt shapes how your bot behaves. Go to Settings → General → System prompt.

A good prompt includes:

• Who the bot is (company name, role)
• What topics it should and shouldn't discuss
• Tone of voice (friendly, professional, concise)
• What to do if it doesn't know something

Example:

You are the support assistant for Acme Plumbing, a Sydney-based plumbing company.
You help customers with booking, pricing, and common plumbing questions.
Be friendly, clear, and concise. If you don't know the answer, say:
"I'm not sure — please call us on 02 9000 0000 for help."

Enable Strict mode

Without strict mode, the bot answers general questions (e.g. "What's the capital of France?") even if your knowledge base doesn't cover them.

Enable Strict mode in Settings → General. With it on:

• The bot only answers questions it can find context for in your knowledge base
• Off-topic questions get a polite decline: "I can only help with questions about [your topic]."
• You can customise the decline message

Add more specific sources

If the bot gives vague answers, the knowledge base might be too broad or lack the right detail. Fix it by:

1. Adding a dedicated FAQ document (plain text source)
2. Breaking a large PDF into topic-specific sections
3. Adding a separate source just for pricing

Check similarity threshold

Pivra uses a 0.4 similarity threshold — if no chunk scores above this, the bot says it doesn't know. This prevents hallucination but can cause the bot to miss valid answers if sources are poorly worded.

Improve this by: rephrasing your source content to match the way customers ask questions.

Test with the Preview

Use the Preview chat (top right on any chatbot page) to test questions before going live. Check:

• Does the bot find the right answer?
• Is the tone appropriate?
• Does it handle edge cases ("I don't know", off-topic) correctly?

Re-train after content changes

If you update a source (edit a web page, update a PDF), go to Knowledge and click Re-train on that source. The old embeddings are replaced with fresh ones.