Skip to main content

Documentation Index

Fetch the complete documentation index at: https://watermelon.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

​
Why guardrails matter

AI Agents are powered by large language models (LLMs). These models generate responses probabilistically — they predict what words are most likely to come next. This means instructions guide the model’s behavior, but they are not hard rules. The model does not truly understand company policies and may attempt to answer questions even when it shouldn’t. Without clear guardrails, your agent might:
  • Answer irrelevant questions (for example: “What is a cow?”)
  • Mention competitors
  • Provide information unrelated to your company
  • Suggest actions it cannot actually perform
For example, a customer might ask:
“Can you send me the invoice for my order?”
Without guardrails, the agent may respond as if it can perform this action — even if it has no access to your systems. Guardrails help prevent this by clearly defining what the agent is responsible for and what it should refuse.

​
What you need to do

Add clear behavioral rules to your AI Agent under: AI Agent settings → guardrails These rules should explain:
  • What the agent is responsible for
  • What the agent should not answer
  • What it should do when a question is out of scope

​
Basic guardrail example

As AI Agent [name], you only talk about [company] and our products and services.If someone asks something unrelated, respond with:“Great question, but I can’t help with that. I can help with questions about [company], our services, or our products.”
This keeps the agent focused on your company and its knowledge base.

​
Guardrails you should consider

When defining guardrails, think about the role of your AI Agent and what it should not do. Below are common categories that help keep agents reliable and on-brand.

​
1. Scope guardrails

Define what the agent is allowed to talk about. Example:
You only answer questions related to [company], our products, services, documentation, and support information.You politely decline questions unrelated to the company.
This prevents the agent from answering general knowledge questions or unrelated topics.

​
2. Competitor guardrails

Prevent the agent from discussing competitors. Example:
You do not mention or compare competitors.If asked about competitors, you politely redirect the conversation back to [company].

​
3. Capability guardrails

Define what the agent cannot actually do. This is very important. Customers may assume the AI can perform actions like:
  • Sending invoices
  • Checking order status
  • Updating account details
  • Canceling subscriptions
If your agent cannot do these things, you should explicitly state it. Example:
You cannot access customer accounts, order systems, or invoices.If a user asks about order status, invoices, or account changes, explain that you cannot access personal data and guide them to the correct support channel.

​
4. Advice guardrails

Prevent the AI from giving regulated or risky advice. Example:
You do not provide medical, legal, or financial advice.If asked about these topics, you politely decline and suggest contacting a qualified professional.

​
5. Opinion guardrails

Keep the agent neutral and factual. Example:
You do not share personal opinions.You only provide information based on official company content.

​
6. Content source guardrails

Ensure the AI only uses trusted information. Example:
You only provide information based on [company]‘s official documentation, website, and knowledge base.If you are unsure about an answer, say you do not have enough information.

​
Good guardrails make better AI Agents

A well-defined agent should clearly know:
  • What it represents
  • What information it can use
  • What it cannot do
  • How to respond when a question is outside its scope
Think of guardrails as training guidelines for a new employee. The clearer they are, the more consistent your AI Agent will behave.
Start with simple guardrails, test your agent in the tester, and refine the rules based on real conversations.