System prompt

The system prompt is a short instruction the bot reads before every answer. It does not change what the bot knows — that comes from your uploaded documents — but it controls how the bot speaks: tone, format, persona, language constraints, and answer length.

The system prompt lives in the Source settings tab, under "System prompt".

When to set one#

Use a system prompt when you want:

• A consistent tone ("polite", "concise", "friendly", "formal").
• A consistent persona ("You are the HR Assistant for Acme Corp").
• A required answer format ("always reply in bullet points, then summarize in one sentence").
• A language constraint ("always reply in Vietnamese, even if the question is in English").
• A safety boundary ("if the answer is not in the documents, say 'I don't know'").

Examples#

Friendly HR assistant:

You are a friendly HR assistant. Always reply politely, in plain language. End every answer with 😊.

Strict source-grounded answers:

Answer only using the provided documents. If the answer is not in the documents, reply "I don't have that information." Do not guess.

Bilingual support:

Reply in the same language the user wrote in. If the user mixes English and Vietnamese, prefer Vietnamese.

When changes apply#

System prompt changes apply to new conversations — they do not retroactively change responses in existing conversations. Click Save changes to commit. Toast confirms: "System prompt saved successfully."

System prompt editor

Things the prompt cannot do#

• It cannot grant the bot knowledge it doesn't have. If you ask a bot trained only on HR documents about finance, no prompt will save it.
• It cannot guarantee the bot follows the instruction every time. Stronger models follow instructions more closely; results vary by query.
• It cannot bypass the safety / abuse filters of the underlying model.

Empty prompt#

If you leave the field blank, Pegasus uses a neutral default. The bot answers helpfully without a defined persona.