#Custom GPTs completely unaware of knowledge base

1 messages · Page 1 of 1 (latest)

gilded iron
#

Bug Report:
Custom GPTs hallucinate and do not recognize the content in the knowledge base. These issues are across all my custom GPTs, they answer without consulting the knowledge base, I have to keep reminding them of the facts. What is the point of having a custom GPT, moreover a well structured GPT with precise instructions and a robust knowledge base if it ignores the prompts, does whatever it wants, responds how it wants and continously apologizes about not addressing hard facts.

Steps to reproduce:

  1. Develop something around X topic (included in the knowledge base)

Expected result:
Exactly what I ask for

Actual result:

Something completely stupid or different with a totally different concept and complete disregard to the base instructions.

Additional information

Browser:
OS:

wild smelt
#

I had this yesterday as well but it seems this morning to be stable again

short comet
gilded iron
short comet
formal fable
#

I experienced this many times too, and OBVIOUSLY the instructions specifies that. In my experience the custom GPT becomes totally crazy if you give more information beyond some kind context or capacity threshold. Before that it works quite good, after that is totally useless, in addition is wasting our time and our usage of the service. For example few minutes ago, asking very simple things to validate before a new task, more or less ok, then I start to give it some information to create a fictional character, after a longer message the custom GPT has completely no idea of which character are we working on, totally unable to extract or analyse the current conversation (much worse than other times and I can say that because I have tried longer conversations other times). The last thing, trying unsuccessfully to make him understand the task, after many useless answers apologizing, now appeared the message that I just reached the usage cap and I have to wait 2 hours... how can that be possible with only 18 messages of mine, I didnt use the service since yesterday. Please addres this bugs bugs because I am sure this is limitating a lot the capabilities of many custom GPTs. BTW all those errored answers due to its incompetence should not be counted for the usage cap, its a waste of time and money

gilded iron
vapid olive
#

from some testing, it seems that documents in its knowledge are treated like a library of information to access on demand, rather than actual knowledge it would have at the start of a conversation. It does seem to do a quick scan of the docs, and it seems to help with this part if the docs have familiar formatting.

I have a use case for which I'd like the GPT to have thoroughly read the entire body of text in each doc (not a huge amount), before its first response, but I've really struggled to achieve this with the custom instructions.

The best result I've had is to format the text as an outline, split it up into shorter docs, and give it custom instructions to summarise the text internally. This is quite unreliable, but what has also helped in combination with this is setting up a conversation starter asking it to tell me about itself. This gives the best result so far, but it's unlikely that it's actually read all the text from beginning to end.

capybarathink Do you have any tips for achieving this, or using another approach? I've tried unsuccessfully to do it by pulling the text from online with an action, instead of uploading docs.

formal fable
#

I tried to send the message I attached as screenshot and it got blocked, I am new to discord, and I don't get why this happened. Regarding its content, I have to say that I became a PLUS user in november, I used the GPT4 and custom GPTs many many hours per week, trying several ways to handle its knowledge base, different file sizes, formats, data structures, using files just for references, ... and never worked as expected, only some moments and with very small amount of data, but lately not even in that case. And I am worried because I think there are bad incentives in this business model, where people pays the same, it does not matter if it works or not, its performance decreases, we invest more time just to start the tasks, we have less usage cap from time to time and me I don't see support of any kind. Its like if you suscribe to a restaurant that starts offering a nice menu, but because its focused in getting more and more costumers, the amount of every dish decreases periodically, and the worse part is that the quality of the food decreases also dramatically. This is the opinion of an AI passionate and ambassador for other people

discord_error.PNG
gilded iron
# formal fable I tried to send the message I attached as screenshot and it got blocked, I am ne...

You nailed it, the Open AI team is selling snake oil at this point. No SLA standards, zero customer service and a completely deficient product that gets worse every week, and even after the countless threads on the same issues from all users they just don't even care to acknowledge it or even respond to anyone on here. I am honestly just waiting on some MSM reporting or some big social media account to come out and expose this for the sham it is. Its a 100B with complete disregard for their paid user base and all they every talk about it their next project, their next capability, the founders wedding... but zero talk on performance and things that matter.

delicate terrace
#

I too have to realize that long form is bad for the bot. It's unfortunate because it takes time cutting everything down to bite-size pieces - like feeding a baby 🙂

#

I wish I didn't have to, but alas

#

also, my custom gpts are mute. i can enter a prompt , but no reply at all.