A week or so, I wrote about my attempt to leverage the virtual environment where ChatGPT creates and stores files to create a "Virtual RAG".

The original post is here:

https://www.reddit.com/r/OpenAI/s/rzMUiF7lD6

I've continued my attempt at this, mostly by attempting to refine the prompt, and have shifted the bulk of the prompt into the custom instructions.

With revisions to the prompt, I've been able to expand the single file to an index of multiple files.

Through the expanded prompt, Chat is reviewing the current conversation, comparing that to files in the index, and retrieving information from the associated file.

I'm sill having issues with it periodically skipping instructions, and also with persistence of the files themselves.

I've managed to extend the persistency somewhat by periodically having chat fully replace the associated file and update the index accordingly.

I've also got it self-determining file names, which has improved accessing the correct files.

Instruction skipping appears to be the largest hurdle.