Hi, I have noticed a very annoying habit of ChatGPT to use only a subset of available data for certain operations (like drawing a time series chart given a table of data). Pretty much all of the time, when the table is past some threshold of rows (maybe 50), it will sample the first few and last few records, and possibly some chunks in between.
This makes the graph totally useless, of course. I can ask it to draw the chart again using all available data, and it complies. But I’d like it to use all the data the first time and in fact every time I ask without specifying otherwise. I tried adding some verbose instructions to the Custom GPT I’m working on, and even explained the reason why it’s important to always use all data points. But it seems to ignore the Instructions when in “I’m drawing a chart!” mode (writing the python code that will draw the chart, and inserting the data into an array at the start).
The file search in ChatGPT is not under your control.
It is not a “habit”; it is how uploaded files “work” - by search retrieval of chunked documents in vector storage, by the AI writing a semantic search query and getting some parts back from all documents that rank.
You can instead demand that the AI never use the file search tool and only process data in Python code interpreter, and only deliver the output as sandbox file links. That: AI can sample the content of a file to ensure it writes good code, but that it shall never attempt to retrieve or act on full data extraction itself due to a limited character count that can be returned as notebook output.
File search and its tool presentation to the AI cannot be turned off when uploading files, another fault. The use of files is co-mingled and confabulated between tools, and even intermixed with user-provided files without distinction.
I thought I’d highlight another challenge you may have in instructing against tools that are placed: GPTs have a lower instruction adherence that previously, are seen as from the user and not an authority, and are readily internally-challenged and disobeyed by both the reasoning and the user input that has equal status.
See: immediately dropping out of persona; equivalence of “GPT” being from a user and not challenged by even “thinking-high”, and an attitude of distrust.
