Skip to main content
The Agent determines the conversation language from the user’s first substantive message and continues using that language. It only switches when the user clearly switches language or explicitly asks it to do so. If your Agent still changes language unexpectedly, another part of its configuration is usually influencing the model.

​
Why can this happen?

The AI model receives more context than just the conversation language. It can also see:
  • Personality and other instructions
  • Example responses
  • Quoted text
  • Knowledge
  • Previous messages
  • Hardcoded responses
If these contain different language signals, the model may occasionally respond in the wrong language.

​
Quick fix

First, make sure one of the main instruction fields, such as Personality, is written in English.
  1. Open your Agent.
  2. Go to AI Agent settings.
  3. Open Personality.
  4. Write the instruction in English.
  5. Save your changes.
  6. Test again in a new conversation.
This does not mean the Agent will start speaking English. Watermelon still determines the conversation language automatically.

​
Check your instructions

​
Avoid unnecessary language rules

You usually do not need instructions such as:
Always respond in Dutch.
or:
Always respond in the same language as the customer.
Watermelon already handles this. Only add a language rule when you deliberately want different behaviour, such as an Agent that must always speak German.

​
Be careful with “say”

An instruction such as:
If a customer wants a refund, say that you can help them with making a request.
may be interpreted as an instruction to use that exact English sentence. If the wording does not need to be exact, describe the behaviour instead:
If a customer wants a refund, explain that you can help them with making a request.
This allows the Agent to communicate the same information in the current conversation language.

​
Check quoted responses and examples

Example responses also act as language signals. For example:
When the customer is upset, say “I’m sorry to hear that.”
If your instructions contain many English examples like this, they may influence the Agent’s responses. Where possible, describe the intended behaviour instead:
Acknowledge the customer’s frustration when appropriate.
The same applies to example conversations. Only use them when they are actually needed.

​
Look for conflicting instructions

Check whether different parts of your Agent tell the model to behave differently. For example:
Always respond in Dutch.
and:
Respond in the same language as the customer.
These instructions conflict. Also search for instructions containing words such as:
  • say
  • respond with
  • reply with
  • translate
  • language
Remove unnecessary language rules and keep instructions as clear and consistent as possible.

​
Can knowledge cause a language switch?

Knowledge can be written in a different language from the customer. This is normally fine. For example, your knowledge can be in English while the customer speaks Dutch. The Agent should still communicate the answer in Dutch. You do not need to translate your knowledge.

​
Does the selected AI model matter?

It can. Different AI models can interpret the same combination of instructions and context slightly differently. Smaller models may also benefit more from short, clear and non-conflicting instructions. However, switching models should not be the first fix. If you experience a language issue:
  1. Check the instructions first.
  2. Remove conflicting or unnecessary language signals.
  3. Test in a new conversation.
  4. If the issue remains, test the same scenario with another model.
If changing only the model fixes the problem, the original model may have had more difficulty resolving the instructions and context.

​
Troubleshooting checklist

  • Test in a new conversation.
  • Make sure Personality or another instruction field is written in English.
  • Remove unnecessary language instructions.
  • Check instructions containing say, reply with or respond with.
  • Check quoted responses and examples.
  • Look for conflicting instructions.
  • Simplify long or mixed-language instructions.
  • If needed, test the same setup with another AI model.

​
For Legacy Agents

If Domain knowledge is visible in the left-hand menu, you are using a Legacy Agent.
  1. Open your Agent.
  2. Navigate to Domain knowledge.
  3. Open Personality, or another instruction tab.
  4. Write the instruction in English.
  5. Save your changes.
  6. Test again in a new conversation.
In most cases, you should not solve language switching by adding more language instructions. Watermelon already handles the conversation language automatically. The goal of troubleshooting is to find instructions or examples that unintentionally interfere with that behaviour.