Hello everyone. When I was testing my assistant api(model gpt3.5 0125), I found out that assistant is not reading all of my file, it is retrieving the file but it’s response is not even slightly about my question. It just grabs first thing I write in my .txt file.
But thing is, assistants was able to retrieve anything I ask with low chance of mistakes couple of days ago. All of the sudden I don’t know what happened.
Other thing is, GPT4 moddel is able to find anything I asked in the file. But I cant use it because it is very expensive for me. I think this is some kind of bug like others we encountering time to time but I don’t really know this might be permanent change too, I have litteraly NO idea.
What do you guys think about it? Do you have the same problem with me and what do you suggesting?
Was anything changed. if anything was changed in your code I would look at that. sometimes it only takes one word or change in instructions or how the data is assembled that creates the issue. I run the GPT3.5 with kruel.ai it does not lie on data, but on things it does not know it trys to guess and states I assume , or you most likely were. based on data. your instructions for the data understanding may need more instruction for steering things. a simple only use the data and do not make things up can go a long way in data instructions. or setting the temp down to 0.3 from 0.5 makes it stay more bang on at the cost of creative language which is what leads to less data focus.
Not really. It is same in both api and playground.
I will try what you said and write the results.
if nothing was changed without understanding how your system works I could not narrow it down. instructions are key. you also have to understand that ai is based on pattern predictions so well 10 results are 100% your 11’th could be a pattern that does not match your current expected thus throws the ai off. and if you are than refeeding your ai back old context with the incorrect info still in it could add to that same pattern and cause the error until its out of your context length. to fix that you need to look deep into your debugs between sessions to find out why it went one way over the other to build up your instructions to handle that odd logic. almost like building a logic process in your instructions 
Noting changed. Main problem here isn’t about prompts or instructions I think. It’s simply only reads the first lines of the file for the reasons I dont know. I tried to say like “Read all of the content inside of files” etc but that didn’t worked neither.
how big is the files? are you use the assistant as the doctreiver? or you just passing the file as text into the message?
Not that much. It is just 92kb and I reduced my files to 1 to work this up again but didn’t worked. I am not passing the text files into the messages I already uploaded them via interface. I tried passing them into the message but that didnt worked neither. I also looked up my threads and copied same instructions that also didn’t worked. I am just hoping that this problem will magicly disappears tomorrow. I am encountered this kind of problems before but after 1 night it was magicly disappeard.
if you uploaded it and called that agent tool and
then on your assistant you just have to have it that the user is asking a message on xyz document find the information along with some additional instructions for what you require. that should work.
id be also checking your api calls with debugs so you can see the raw inputs and outputs from the call to make sure there is not some error or context length issues where it error than the ai will just make something up depending on your logic to handle errors.
1 Like