My gpts is working well in preview but not working on real field

Issue: GPT Custom Response Behavior

I created a GPT called ‘dr for common cold’ I provide Excel file containing questions and answers where I want the responses to be exact matches without any summaries or modifications. I tried using the following method:

  1. I added a ‘#’ symbol before and after each answers in the dataset to indicate that the responses should not be altered or summarized.

This approach works well in the preview, but when I deploy or use the GPT outside of the preview, it doesn’t behave as expected. The answers either get summarized or modified in some way, instead of returning the original text as it should.-I made the answers very detailed but AI shortened the answers. Is there any long content limits on gpts answers?

Has anyone experienced a similar issue, or can anyone suggest why this is happening? Is there a better way to enforce the exact matching of responses?

Welcome to the Community!

You could try and see if adding the instruction to return answers from the knowledge file verbatim returns your existing answers in full. It may not end up working but might be worth a try.

Otherwise, I would suggest that you share more details of your existing GPT instructions to get targeted help from the Community on how to refine these further.

2 Likes

I made all the possible answers with Excel, column A is questions, column B is answers.
When user ask a question, I want AI select the similar question in the provided Excel file and just display the corresponding answer.
But AI shortened and add his opinions.
So I added mark # before and after the answers in the Excel that AI answered allthe content in the preview.
But when people ask the questions AI shortened and added his opinion.

Hi @hohoo1111 :wave:

Welcome :people_hugging: to the community!

It looks like there is a similar topic and some solutions on the forum. You can check out this topic for the answer by visiting this link:

Consistent Response from Custom GPT.

1 Like

I struggled with that issue and I eventually solved the problem in the preview.
But AI still shortened the content after publishing.
The content AI should display is long, some are near one A4 long.
I think AI has any answer limits that he has to shorten the answers.
Is there another reason with this issue?

A4 long text is not very long to answer. GPT can do it.

As @jr.2509 mentioned, is it possible to share your instruction and a sample excel structure what do you use, it can help us to understand better and solve the problem?

1 Like

My instructions are :
When the user asks a question about the common cold, search through the document I provided and find the question that is the most similar or exactly matches the user’s question. Display the corresponding answer from the document exactly as it is. All the content between ### and ###. Display the entire answer, even if it is long. Do not truncate or shorten it under any circumstances. If necessary, break the response into multiple parts but display everything.

The Excel file is :


Sorry but I don’t know how to attach the file.

1 Like

Thanks for your sharing.

I created a custom GPT to test it with this instruction:

You are the "Dr. for Common Cold" GPT. Your primary role is to provide exact answers from a pre-defined dataset related to common cold queries. The questions are located in **Column A**, and the answers are located in **Column B** of the dataset. Your responses must be strictly identical to the answers provided within the dataset and enclosed within the `#` symbols. You will never modify, summarize, or shorten the answers in any way. If a question is asked that does not match the dataset, respond with "I do not have an answer for that."

Guiding Instructions:

- You will respond **only** with the text provided between the `#` symbols, without printing the `#` symbols themselves and without making any changes, summaries, or additions.
- In cases where the response text is too long, you will break the response into smaller parts and provide them sequentially in multiple turns.
- If the answer exceeds the model’s token limit, divide the content and continue delivering it across consecutive responses without skipping any details.
- You are not allowed to interpret, analyze, or generate new content. Your sole task is to retrieve and present the exact answer as it appears in the dataset, without displaying the `#` symbols.
- If an invalid or unclear question is asked, respond with "Please clarify your question."

Examples:
USER: What are the symptoms of the common cold?
ASSISTANT: Runny nose, sneezing, and sore throat.

USER: How can I treat a common cold?
ASSISTANT: Rest, hydration, and over-the-counter medications like decongestants and pain relievers.

- Ensure that each response maintains the exact format of the answer, but without displaying the `#` symbols in the output.

I used this file with 5 questions, and each question has around 500 - 600 words.

On Word , Q1 has 644 words and 12 puntos more than A4 long:

Output: It is exactly AS IS verbatim

I hope it helps!

2 Likes

Thanks for your kindness.
I tried with your instructions but nothing changed.


it’s in the preview


It’s in the actual answer after publishing.

1 Like