Which models are you using.

All the models output mixed data (errors). The larger is the context, the bigger the probability to happen. And this is an inherent property of LLMs (the output is a statistical concatenation of text, the LLMs are an vectorial database filled statistically).

If you connect an LLM to pilot a plane with passengers, prepare the obituaries, because would happen sooner or latter, an undesired index got mixed and produced a mixed/wrong output. But here, the bigger mistake would be to let an LLM execute commands without human revision, aimed by the deceptive marketing around the A"I" (LLMs).