Extend the notion of Chat Message and Chat History

[InAnYan]

An LLM (chat model) is limited only to system message, user and AI messages.

But the real notion of the chat message is more complex:

1. There are thinking stages of AI response. In API they are AiMessages but it should not be seen as AI answer.
2. AI answers contain references. But these references are only mention in UserMessage.
3. There are various kind of "status" messages: model or something else was changed, error messages, auto retry mechanism. Those are not AI messages and not user message (and not system message).

To sum up, we just need to create our own hierexhy and message types. And then develop a visitor that will convert JabRefMessage to dev.langchain.ChatMessage

[ThiloteE]

In theory, we have something like input > output.

Input: System message, user message
Output: Response by the model.

Model makers mostly use jinja as chat template.

See https://huggingface.co/Qwen/Qwen3-8B/blob/main/tokenizer_config.json#L230
See https://github.com/google/minja (which is used by llama.cpp)

I suppose this template format is not propagated through an API?

In any case ... this issue here is very complicated, because every model has different responses and different chat templates, so it takes a lot of work to make JabRef compatible with every model. Minja does just that and aims to be compatible with the vast majority of models and you can see how much work that is.

[InAnYan]

I didn't mean LLM messages. More UI messages

[ThiloteE]

You are right though, a first start would indeed be to address <think> </think> in the UI.

[InAnYan]

Did I really send 2 identical messages in this issue?