GPT generation follows a hidden email pattern

Prompting
1

I am working with gpt-3.5-turbo-16k model to generate email content.
I prompted Gpt to use ‘Dear User’ to start, but it seems there is an implicit template of the form ‘Dear User, Thank you…’
What I need is for it to start with ‘Dear User’ and not Thank You.
Does anyone have any suggestions for this problem?

2

Hi,

Have you tried adding a one-shot example in your prompt + reducing the temperature?

3

Hi @mbassi
I set the temperature to 0, added an example and said GPT no Thank you but it still doesn’t work.

4

@TuanAnh ok got it, is if ok for you to share the system prompt?

5

This is my system prompt

{
	role: 'system'
	content: `Generated Reply: Generated Reply must be based on the following description: \"ask for lease intention.\". The reply should have paragraphs and linebreaks to be as clear and legible as possible.      Your email signature should be: '\\n Kind regards,\\nDavid\\nProperty Manager'      The response should be in a Professional tone and must take into account the conversation history as well as all other relevant data provided.      You must start with 'Dear User'.      You NEVER say Thank you to anything.\n      Example reply: 'Dear User,\\n\\n We will follow our internal process to deal with the matter. \\n Kindly let us know if you need further assistance.'\n`
}
6

Ok, few things you may try:

  1. Rule Clarification : It might be beneficial to define our rules more precisely. Clear and distinct guidelines will help in generating more accurate and relevant automated responses, reducing misunderstandings.

  2. Prompt Optimization : I’ve conducted some tests with the prompt, setting of 0 temperature and 3.5 as a model. It seems to resolve the issues with the ‘Dear User’ and ‘Thank you’ phrases. You need to adapt it to your case

  3. as a bonus: If the signature is always the same, maybe you could add it programmatically to save some bucks

You are a chatbot that generates automatic email replies.

Rules:
- The reply must be based on the following description: \"ask for lease intention.\". 
- The reply should have paragraphs and linebreaks to be as clear and legible as possible. 
- Do not insert any signature at the end.
- The response should always be in a Professional tone and must take into account the conversation history as well as all other relevant data provided.     
- You must start with 'Dear User'.  You should always avoid "Thank you for your question" or similar behaviour.\n      

Example:
User: "I need XXX"
 Agent: """Dear User,
We will follow our internal process to deal with the matter. 
Kindly let us know if you need further assistance."""
7

Thanks @mbassi
I tried your way but it seems GPT is not up to my expectations.
There’s still a ‘thank you’, do you know any other way to fix it?

8

Subject: Re: AI Model Writing Issue

Dear Colleague,

I hope this email finds you well. I’m writing in response to your question regarding the issue with our AI model, specifically its tendency to continually write “thank you” at the beginning of its outputs.

The most effective strategy to prevent the AI from generating these overtrained form-letter-like responses, particularly for uses such as “email” or “summary”, is to avoid any reference to these output formats in our instructions to the model. Instead, our focus should be on the composition and form of the text itself.

By doing so, we can guide the AI to produce the desired output without it falling into the pattern of starting every text with a “thank you” or any other repetitive phrase. This approach should help us to better control the AI’s output and ensure it aligns more closely with our requirements.

I hope this provides a clear solution to the issue you’ve been experiencing. Please feel free to reach out if you have any further questions or need more clarification.

Best regards

2 Likes
9

Hi @TuanAnh, , I cannot recreate the same error you are experiencing with my solution.

Also, @_j solution works fine (it’s cleaner and will save you some more bucks). I think it could be structured something like this:

You are a chatbot that generates automatic email replies.

Rules:
- Any eventual rules you need to enforce

Email Structure:
- "Dear user," + <corpus> + "Best regards"

Example:
User: "I need XXX"
 Agent: """Dear User,
We will follow our internal process to deal with the matter. 
Kindly let us know if you need further assistance.
Best regards."""
1 Like