How to customize the template for chat_ history

[zyg992018]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

history = ChatMessageHistory()
    llm = LLMRemoteChat(url="")

    prompt = ChatPromptTemplate.from_messages([
        ("system", "you are a student."),
        ("placeholder", "{chat_history}"),
        ("user", "{input}")
    ])
    llm = prompt | llm
    chat = RunnableWithMessageHistory(
        llm,
        get_session_history=lambda session_id: history,
        input_message_key="input",
        history_messages_key="chat_history"
    )
    chat.invoke(
        {"input": "hello"},
        {"configurable": {"session_id": "unused"}} 
    )

I noticed that when I manage my history, chat_history is automatically converted into the following format:

System: you are a student.
Human: hello

However, this format is not what I want. For example, in the transformers library, I can use tokenizer.apply_chat_template to add special tokens to keep it consistent with the training. But I cannot reformat the chat_history template from the code because it has already been converted to the default format in the LLM's _call method's prompt parameter. I also don't know how to access the history in the LLM's _call method or how to rewrite RunnableWithMessageHistory to format historical messages, as the RunnableWithMessageHistory class doesn't seem to have an interface for this purpose.

Motivation

empty

Proposal (If applicable)

No response