Our goal is to give you the agent that can be a reliable and trustworthy agent in a business environment - out of the box.
We have made considerable efforts to make our agents much less susceptible to abuse and to stay on topic, do not hallucinate, and avoid engaging in sensitive or explicit topics (politics, philosophy, NSFW, etc.)
By default, the agent is meant to provide concrete, on-point answers and admit if it does not know the answer.
Some website visitors like to abuse agents and essentially use them as "free ChatGPT," which can cost you money if you install it on your website!
What are Guardrails
This concept of guiding the agent towards more desirable (and away from less desirable) behaviors is called “Guardrails”.
In Chatwith, the Guardrails are implemented using prompt engineering: the practice of communicating desired Large Language Model (LLM) outputs using natural language.
While having Guardrails in place is useful, sometimes you may want to create an agent without them - and imprint your own rules.
How to change the Guardrails
You can adjust the Guardrails level in your dashboard ‘Settings’ section here.
You can choose between three levels:
None | Low | Max (default) |
✗ No guardrails, AI can answer outside of context and off-topic
✗ No requirement to answer in the same language or with formatting
✗ No context on Actions except for configuration details
✓ Full freedom to customize the base prompt as needed | ✓ Basic guardrails to stay on topic, weakly enforced
✓ Encouraged to answer in the same language and with formatting
✗ Might make mistakes when using Actions
✓ Flexibility to enhance the base prompt | ✓ Avoid off-topic, sensitive or explicit subjects, assertively decline to answer
✓ Must stick to instructions, answer in the same language and with formatting
✓ Most reliable when using Actions
✗ Customizing or overriding the guardrails is difficult |
Be cautious when changing the Guardrails level. It significantly affects the overall function of your agent and can have unexpected consequences.
Tips
- Make sure to thoroughly test your agent after changing the Guardrails level.
- Ensure that Actions (if configured) work correctly after changing the Guardrails level.
- Consider adding your own guardrails if using an agent on None/Low Guardrails levels. You can use the System Prompt field.