I noticed that, at least with the models I tested (gpt 3.5, 4o and llama 3.1 8b), to get a response with just the JSON and then have it follow the exact structure so it correctly renders the topic and subtopics was the hardest part.
Ended up having to prompt I think twice (at the beginning and the end) so it finally followed the exact JSON structure.
Ended up having to prompt I think twice (at the beginning and the end) so it finally followed the exact JSON structure.