Fine-tuning and different system messages in the conversation

I have a question about how to send training data knowing that my system message changes multiple times in the same conversation, but for training, you can only send one, for example:

System Message 1:
You are the chatbot from Rotulados Jas and you are assisting customer Luizhangela Melo.
-Friday 01:09 PM.

user: Hello, good morning. I would like to place an order for glossy white vinyl. 1.50m
assistant", “content”: “Hello Luizhangela, good day!”, “function_call”: {“name”: “search_product”, “arguments”: “{"query":"adhesive white vinyl"}
function”, “name”: “search_product”, “content”: “244, Adhesive White Vinyl 137cm, $4|441, Matte Adhesive White Vinyl 137cm, $4
assistant”, “function_call”: {“name”: “shopping_cart”, “arguments”: “{"id":2,"quantity":1.5}”}}
function", “name”: “shopping_cart”, “content”: "1.5 of product 2

System Message 2:
You are the chatbot from Rotulados Jas and you are assisting customer Luizhangela Melo.
-Friday 01:09 PM.
SHOPPING CART:
1.5, Flat White Vinyl 61cm, $3.75
Total: $3.75 or 134.2125bs

assistant: Okay, this would be it:
1.5, Flat White Vinyl 61cm, $3.75
Total: $3.75 or 134.2125bs

user: Perfect

So far, I have only come up with two options:

  • Divide the conversation when the system message changes. Problem is, I lose context for training.
  • Send the shopping cart in the function response. Problem is, it would generate a lot of tokens that would later be irrelevant, among other issues.