how come the accumulate response

Bug?

seems like it might break on an chat model or am I a silly goose

it would get converted into a single human message at some point

the selector is more for selecting an engineered prompt for the right model (chat vs completion)

seems like it almost does same thing since it calls format_messages in the llm predictor

guess its missing the system call?

Yea exactly

should I fix that

Go for it!

if there is no system call

what does that actually effect?

well considering it has this in the system prompt; "1. Never directly reference the given context in your answer.\n" that's important

Yea the system prompt is just a type of message that the LLM pays attention to

usually its for "non-conversational" instructions