Rebuilt ChatSession class

[philippjbauer]

• Saves with serialized ChatHistory of session
• Only allows use of ChatHistory.Message (instead of raw text) for easy post-processing with IHistoryTransform implementation
• Provides History Management methods
• Allows user to regenerate last assistant message

[martindevans]

Thanks for re-submitting this based on master! Overall looks good, just that conflict to resolve and one comment/question.

[philippjbauer]

Thanks for re-submitting this based on master! Overall looks good, just that conflict to resolve and one comment/question.

I don't have write access to the file to resolve the conflict.

The lines in red need to be removed, the ones in green kept:

[martindevans]

I don't have write access to the file to resolve the conflict.

I've resolved that for you, please check that I did it right!

[philippjbauer]

I don't have write access to the file to resolve the conflict.

I've resolved that for you, please check that I did it right!

LGTM 👍

[martindevans]

The tests seems to be failing due to a type checking error, looks like something isn't right with that merge?

[AsakusaRinne]

This format of prompt looks good. Is it a commonly used format in LLM now?

[philippjbauer]

The tests seems to be failing due to a type checking error, looks like something isn't right with that merge?

Hmm, I didn't run into these issues before. I thought I had updated everything that depended on the ChatSession class before pushing this PR. I'll have a look tomorrow and fix them up.

[philippjbauer]

This format of prompt looks good. Is it a commonly used format in LLM now?

What exactly are you referring to? I haven't touched the default IHistoryTransform implementation so I'm not sure where you're referring to the prompt format.

[AsakusaRinne]

What exactly are you referring to? I haven't touched the default IHistoryTransform implementation so I'm not sure where you're referring to the prompt format.

I mean the following format in the example prompt:

{
    "messages": [
        {
            "author_role": "System",
            "content": "Transcript of a dialog, where the User interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision."
        },
        {
            "author_role": "User",
            "content": "Hello, Bob."
        },
        {
            "author_role": "Assistant",
            "content": "Hello. How may I help you today?"
        },
        {
            "author_role": "User",
            "content": "Please tell me the largest city in Europe."
        },
        {
            "author_role": "Assistant",
            "content": "Sure. The largest city in Europe is Istanbul, Turkey."
        }
    ]
}

[philippjbauer]

What exactly are you referring to? I haven't touched the default IHistoryTransform implementation so I'm not sure where you're referring to the prompt format.

I mean the following format in the example prompt:

{
    "messages": [
        {
            "author_role": "System",
            "content": "Transcript of a dialog, where the User interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision."
        },
       { ... }
    ]
}

This is simply a json representation of the chat history. This will be ultimately transformed by the IHistoryTransform.HistoryToText() method into the prompt format that the user specifies for the model they're using. Saving it as json makes the system "prompt format agnostic".

As such the user (developer) is able to introduce multiple implementations of the IHistoryTransform interface for multiple models and easily switch between them. I ran into the problem where I had to use regular expressions to extract the content for each turn from an already formatted chat history (like chat-with-bob.txt).

This also allows us to package specific IHistoryTransform implementations for popular prompt formats, like ChatML, Alpaca, etc. The DefaultHistoryTransform is kinda-sorta working for most, but not everyone is aware that the correct template does make difference for most models.

[philippjbauer]

The tests seems to be failing due to a type checking error, looks like something isn't right with that merge?

I synced my master with the latest to get the files that caused issues and updated them to use the new ChatSession integration. The sample for the Chinese character encoding works, but I have a hard time evaluating whether its output is actually correct (I sadly don't know Mandarin). This was a best guess effort on my part.

@AsakusaRinne I saw that you committed the code, are you able to verify this?

[AsakusaRinne]

@philippjbauer I've added some changes and the Chinese example is generating correct answer now. It's absolutely a good work!

[philippjbauer]

@philippjbauer I've added some changes and the Chinese example is generating correct answer now. It's absolutely a good work!

Thank you!

[martindevans]

Thanks for fixing the merge conflict and the Chinese example. This all looks good to me, ready for merging?

[philippjbauer]

Good on my part

[martindevans]

Thanks for all your hard work on this PR!