Problem

Often GPT returns invalid json. So far in such a situation we request each segment separately. This leads to missing context for GPT and therefore lower translation quality.

Sometimes even this fails.

Please see linked TS issue for a perfect example task.

Solution

Find out how to tell GPT in the prompt, what has been invalid in its previous answer and ask it to send the answer again but with corrected syntax.

If this fails for 3 times, throw a warning and now request segment by segment.