What will soon be a dumb question (just as soon as I know the answer), why in the agentchat_RetrieveChat is the userproxy doing the rag lookup and not the assistant? The assistant in the contrib does nothing. I thought that the userproxy acted on behalf (actively or passively) of the human, and the assistant(s) do the LLM heavy lifting, but it seems reversed in this example. Does it matter? Why is this the preferred way? Thanks!
#RetrieveChat assistant vs. ragproxyagent
feral lodge
paper acorn
I do think the naming and the design could be done differently. @magic slate for your awareness.
magic slate
Hi @feral lodge , thanks for question. Assistant only calls LLM, but what should be put into the context is a decision of the human user, that's why it was designed like this. The assistant needs to handle some special command from human input, currently it's "update context".
@paper acorn , do you remember we discussed about removing rag assistant, just using the default assistant? It turns out that we can't handle user's input in retrieve_user_poxy_agent without modifying check_termination_and_human_reply. So it seems better to simply add a function to retrieve_assistant_agent.
paper acorn
One alternative is to handle the reset of the assistant agent inside the proxy agent