Different answers of endpoint /completion using curl and using chatui

[andreys42]

I wonder what can cause difference in Llama3-8b answers for the same promt when use curl to llamacpp http server endpoint vs when use chat ui requesting the same llamacpp server.
curl respond with messy output, looks like the model ignores rules in system_prompt parameter, while chatui case looks good to me.

My guess is that reason is chatPromptTemplate which I use as parameter in chatui config vs raw string as prompt in curl request. Can this cause differences. If so, how can I set chat template when use /completion endpoint
Thank you for any suggestions