We’re trying to use assistants for data analysis. It’s a new assistant each time since the KB always changes.
The new assistant always fails the initial questions about the KB unless simple. Invariably, after goading it about its mistakes and prodding it to follow its instructions to use retrieval tool and code interpreter … it will work.
What tricks are you using to get newly created assistants to fully index the KB and load into memory? We’re failing to replicate the experience with a custom GPT where you upload the data set into the prompt where it analyzes then does an incredible job.