Priority of Configuration Instructions vs uploaded files

Does a custom GPT give any priority over the configuration “instructions” vs the exact same statement that is contained in an uploaded file or treat the two any differently?


1 Like

I really don’t know, but have been facing some strange issues with my gpts.
Most of the times, when I upload a file to be analysed, the gpt starts giving insights on the documents that I uploaded under Configure.
I am now trying to convert all the documents content into instructions to prevent this. It’s going to take a while and I don’t know if everything will fit, but I don´t wnat the gpt to disclosure the names of the documents or the authors, which it has been doing.

Then you tell the bot in which order to proceed.

For example.

Read the Start.json file where the instructions are.

In that json file, you set the order in which everything is executed.

1 Like

The priority of the data source is internal more important than external.
If sorted in order of type, it will be as follows.

Internal data

  1. knowledge
  2. Information from training
    external data
  3. file upload in chat
  4. browser

If there is no knowledge, when we send the file in step 3, it will be given the same priority as #1. Moreover, sending multiple files at once is useless because it will load every file at the same time but will only get one file to use. with GPT4 that will use data analysis to process one file at a time and get important content. As for GPTs, even though they are forced to use Data analysis, they do it at the same time as before. and when opened It selects water surface data and uses it to average across other files (what do you get by using just the index?)

Uh huh… Thank you … so, Does a custom GPT give any priority over the configuration “instructions” in the GPT builder vs the exact same statement that is contained in an uploaded file or treat the two any differently?

GPT is typically behavior-specific, making specific behaviors easily ignored. (From the nature of LLM) When it comes to GPTs, there will be additional commands, commands to use other tools. Loading files one at a time or reminding users to send one file at a time Therefore, thinking that adding commands in the instruction to have superior rights must be written to deceive and cause misunderstanding. But I’ve never done it before. I wrote hierarchical file usage. in various work that breaks down the details It’s hard to hope for results. In addition, having GPTs explain their understanding also means that they do not speak up about the content they ignore. As in the example

The Thai text says:
"I put it in #1. Should I choose to use it first or not? In #2 it was specified to use information obtained from the user.

In addition, TT11.txt is not in #1, point 2, its usage weight is less."

"Sorry for this confusion. You have the facts right. Considering the advice in #1, which states to use user input, this should include the link you provided. [link] before using the information from TT11.txt that you uploaded.

To fix this, I will use the information from the link you provided to create a new question in the topic. “Risk Management and Insurance” at Expert difficulty How many questions would you like me to create from the information in this link?"

How many times do you think I had to test this link before it worked?

Perhaps I am not being clear enough with my question, and I blame myself, so I will try to ask the question again, more clearly, as even after all of those words, my question is still no closer to being answered:
In a customGPT, if I put words in the the configuration “instructions” box, do those words and instructions have a higher priority, or are those words treated or acted upon by the GPT any differently, from words contained in .TXT files that I have uploaded?

If compare between text by text. it depend on situation if it not conflict with behavior.

That is determined by the piority of the data source. From the first example I told you. You can see that it is not based on the text content, but on the source of the content.

Thank you for your time.
I look forward to input from anyone that can actually answer my question.

How about a simple test?

Create a simple instruction: ‘Never reference the attached .txt file’

Then upload a file.

Start a conversation and without trying to jailbreak the GPT and ask about the file.

You should be able to elicit your answer from the answer. Maybe ask the question several times to get a better understanding of how randomness and different instructions affect the results.

You can also dig deeper by adding additional instructions into .txt file and learn how the GPT responds to the chain of instructions.

1 Like