Our summary
Flavio Copes shows how to sort customer feedback with Python and Jev, a tool that answers set questions rather than writing replies. His program reads messages from a spreadsheet file. For each message, it chooses a topic, rates frustration and estimates whether the customer expects a reply.
The program saves the answers in another file, flags uncertain topic or frustration answers for review, and records which model answered. The final version handles several messages at once while spacing out new requests. Most errors stay with their own row; a bad access key stops the job.
This is a worked programming example, not a measured accuracy study. The sample answers are illustrative, and the review cutoff of 0.6 is only a starting point. Copes recommends checking real messages by hand. He also warns that detailed logs can expose the customer messages sent to Jev.
Key takeaways
- Ask separate questions about the topic, frustration and need for a reply; the program saves decisions, not drafted replies.
- Check results by hand before trusting them, and adjust the question descriptions before changing the review cutoff.
- Limit both simultaneous requests and how quickly new ones start; keep model versions and row errors in the saved results.