add add_system_message method to ChatMessageHistory

[mziru]

I added an add_system_message method to ChatMessageHistory (the default implementation of BaseChatMessageHistory). Should pass linting this time!

Seems useful e.g. to be able to initialize a conversational agent with a system message in memory.

Though is there a reason I'm not thinking of that this wasn't already implemented, i.e. to go along with add_user_message and add_ai_message? (Even if not all models have this feature)

If not, should this be added as an abstract method to the base class and implemented elsewhere, e.g. for CosmosDB and DynamoDB?

[hwchase17]

completely agree it can be useful to have a system message in the prompt, but thats more the role of the prompt itself rather than the chat history

[mziru]

completely agree it can be useful to have a system message in the prompt, but thats more the role of the prompt itself rather than the chat history

yes that definitely makes sense. i was running into validation errors specifically around using system messages with the chat-conversational-react-description agent, which expects a memory component. was able to resolve these by adding the system message to the memory before instantiating the conversation agent. but i'll take a closer look to see if i can resolve these with the prompt itself rather than messing with the chat history (and i'll follow up if not).

thanks for the feedback.

[baskaryan]

sounds like this isn't exactly the solution we want, closing for now but let me know if I'm missing something!

[thejollyrogers]

I have found it useful to append a system message to the end of the chat history for gpt-3.5-turbo which even after OpenAI's upgrade still has trouble paying attention to the system message sometimes. For example I have this line after I save_content with the ConversationBufferWindowMemory

memory.chat_memory.add_message(SystemMessage(content="Please format your output in json format with the following fields...."))

I find this gets 3.5 to be more reliable, even though I'm including the same directive in my prompt.