Uploaded custom GPT knowledge files getting truncated?

[nvpc2001]

I’ve been trying to upload relatively short markdown files into my custom GPT knowledge and they appear to have been shortened.

Example:

1. I uploaded a 1700 character brand strategy markdown file in the knowledge
2. Prompted the custom GPT to recite it verbatim
3. GPT only returned about half of the file and told me that the file might have been truncated
4. I tried uploading and asking for recitation in normal chatGPT environment and got the same result

Is anyone else experiencing the same issue? Is this a temporary error or a permanent limitation?

[polepole]

Hı, welcome to the community!

If you make enable Code Interpreter & Data Analysis, you can get better result.

And try this prompt:

Recite the full content file exactly as uploaded. Preserve all formatting, including headings, lists, and code blocks, without truncation.

I tried, and it provided full a sample md file content with 1996 character, without missing any character.

[nvpc2001]

Hi.

I’ve always had Code Interpreter enabled. Tried your exact prompt and still got the same result.

What’s the “unofficial” limitation to be expected for custom GPT knowledge file?

I was hoping that a custom GPT would be able to “know” all my essential business context. Having only 20 files with 2-3 pages each is absurdly limited. What’s even the point of 512MB allowance per file?

Can I have larger knowledge base for custom GPT if I link it with external sources like Google drive or Notion?

[_j]

Files, actually a file search tool the AI can call, is powered by a vector store embeddings search engine. The extracted text is chunked, and has search metadata.

The AI writes a search query, and receives the top-rank results, potentially out-of-order.

There is also some of the partial document text placed in the AI context as preliminary information. ChatGPT may be reciting just from this and not invoking a search, whereas a GPT could be given instructions to ensure that file_search is mandatory and contains knowledge required to answer.

Code interpreter would take even more instructional goading to write python scripts to return the contents of files from its mount point, which also have a limited character return.

The subscription level of the GPT user, and their “context limit”, will determine how much input knowledge can be placed, which can make a massive difference between an 8k “free” and a 128k “pro”.

[nvpc2001]

I have Plus.

I tried uploading the same files in normal ChatGPT environment and it can parse them just fine. Even my longest file with 35069 characters and 808 lines.

Is the context limits somehow smaller for custom GPTs?

[_j]

@nvpc2001

Is the context limits somehow smaller for custom GPTs?

It is possible that OpenAI unified the experience of GPTs so that you can see the results a free user would get (as “advertising”).

…along with pushing a model only worthy of being free on GPTs, from the original gpt-4-turbo-powered GPTs at release. Making them now a novelty.

Try the suggestion of forcing the AI via instructions to use file_search tool, or just ask for it as an experiment, and see the document. You can just put a new fact at the end of document text needing the last chunk or complete observability. Then you would not be asking for reproduction that may cut off prematurely when the AI doesn’t like to write long responses.

[nvpc2001]

Thank You.

I tried “use file_search tool” in both chat and custom instruction and asking the last section/sentence of the document. The GPT still only returned limited portion.

I think you’re right about Custom GPT being throttled down to free user level. Kinda sad. I decided to “migrate” everything into a project and it’s working as intended now.

[nvpc2001]

Update: Hoping that openAI sees this.

I created another custom GPT and it can parse the knowledge files just fine. Looks like the file truncation issue was only affecting certain custom GPT and most definitely a bug.