Even if it's just text, the point is that LLMs are currently still really bad. Some models we still have to correct it for outputting Markdown code fences for

Even if it's just text, the point is that LLMs are currently still really bad. Some models we still have to correct it for outputting Markdown code fences for JSON (we just brute force and string replace it). We have to give it minute details or fine-tune it on a large enough sample set to get the results we want. Even just LLMs right now still have a ways to go in terms of how good they are at working without needing precise instructions and micro-corrections and throughput.