Use alternating Human / AI messages in M4, Assistant

[rlancemartin]

@robbiemu noted

In module 4, research assistant, some models will not respond if the final message in the conversation is an AIMessage (which is always the case), so you may have to modify search_wikipedia and search_web to "humanize" the last message like:
msg = HumanMessage(content=msg.content). This is true for local models but may even be true for non gpt models (When I run mistral-nemo locally, it will not respond. If you have a ChatMistral, that would behave the same way). Likewise the quality will be a lot lower if we do not do the same for generate_answer, as its references will be intact but no write up will be done by the expert

[rlancemartin]

@robbiemu do you mind trying this?

#34

[robbiemu]

Sure, I am just about to do that .. just FYI, here is what I had actually done when working through the module: https://gist.github.com/robbiemu/2704b22215694f4e5539ae601178d922