Are you finding this to still be an issue? I last updated my custom gpt I back in late feb or earlier march as best as I could tell. Until I pushed several updates yesterday. I had not noticed any issues until I updated some of my docs yesterday and went to test its knowledge, but maybe it was just pulling from memory.
I have 4 main .md docs 230k characters or less each. and 1 json file with less than 100k characters for a combined total of < 700kb.
What I am finding is the custom gpt is on the initial read of the document it is only indexing about 1/10th of the document (139 lines out of 1151 lines for example give or take), then it proceeds to act as if the other information does not exist at all. I know which line, as It will tell me the last thing it can see in the document, Until I ask something like how many characters does a specific document has, and then it may change the last line of what it sees. For me knowing what is in the document and referencing it past the initial point will not in most cases get the custom gpt to discover the information.
Also I know it can see the headers as it will tell me all of them past the point of information it initially tells me using the prompt “let’s do some data validation. first provide a breakdown of the document by top-level sections (# H1 headers), including the number of characters, words, and lines for each section. Then read, index and analyze each section one after the other until the end and commit it to memory. Then I will test your knowledge. For this validation we will use facts only, no speculation, inference, postulation, proposals, or similar. for now, only check xyz.md.” it lists them out then proceeds to tell me nothing exists.
I have spent a day and a half reformatting my bookstack docs, and re-exporting to make no progress. They are very simple at this point H1-h4, and some 1 -2 levels of bullets and a small amount of Unicode. Nearly non of which is at most of the initial stopping points in the document. The .mds are dialog, and informational driven, and the one .json is an export from a database with minimal locational data.
The gpts theorize I am hitting a token limit in the document(s) but I have received conflicting info on the current limits. They also tell me I am limited to 20 docs, which splitting 1 into the tiny chunks its reading would be half that limit. So I am at my wits end with this thing, having wasted nearly two days trying to figure out why it wont interact with knowledge anymore.
If it cant read these docs, it is of no use to me… how are any of them functioning anymore? I am on GPT plus btw.
Anyone have any ideas or workarounds?
Sorry for the long post!
