Hi,
I’m working on making a blog generator using API v2 and during testing, I made a multiple assistants, deleted them, made major changes to instructions, and I finally have an assistant (9th) that works more or less the way I want.
I always delete an assistant when it’s no longer in use and for each blog, I submit the input text, (no files or vector spaces), create a new thread, get the assistant output as json_only and then delete the thread.
However, in 300 blogs generated there have been ~10 cases where the latest assistant has responded based on a mix of instructions of all previously assistants I made and generated. I don’t understand why but the (troublesome) behavior is heavily skewed towards the instructions I gave to the first ever assistant in this project. Almost feels like a ghost chasing my project.
Here are some way I minimized this behavior, which each new version of assistant, I had to add a lot more of ‘new’ instructions, at the stage of creation, in the message, in the json response description. I also found that the biggest way to minimize this behavior is to use a completely new assistant name and re-phrased instructions.
I don’t have the code to share as I was iterating through a long script multiple times but this behavior makes me very concerned about potentail guard rails between different assistants.