If this is a non-realtime process on an existing data set, you might consider using OpenAI’s own batch processing - you send a file of chat completions requests in API JSONL format for a 50% discount, a 24 hour turnaround for a file with responses (or recently for -mini, a 24 hour “wasn’t run” often).
Without even knowing what language you are using, it is hard to guess at any faults in techniques (although fault one is a programmer calling the API “ChatGPT”).