Has anyone come across this error when using multiple agents and groupchat? "GroupChat select_speaker failed to resolve the next speaker's name. This is because the speaker selection OAI call returned:"
#GroupChat "Next Speaker" Issue
vivid hound
unkempt swallow
This is likely due to how the next speaker is selected.
There are some options here.
- speaker_selection_method: the method for selecting the next speaker. Default is "auto".
Could be any of the following (case insensitive), will raise ValueError if not recognized:
- "auto": the next speaker is selected automatically by LLM.
- "manual": the next speaker is selected manually by user input.
- "random": the next speaker is selected randomly.
- "round_robin": the next speaker is selected in a round robin fashion, i.e., iterating in the same order as provided in `agents`.
The error you have suggests that you are using an LLM to determine the next speaker. Now, less capable LLMs may not return a valid name of the next speaker ... leading to the error. One thing you can try is try other deterministic methods such as round_robin etc and verify performance for your task.
@cosmic nacelle might be able to provide additional perspectives here.
cosmic nacelle
Yeah there are many situations where that LLM call can fail at present. There are a few prs in progress to help address that. In the meantime use round_robin, or set allow_speaker_repeat to false (I think that’s what the option is called). You should also set last_n_messages to “auto” if using code execution
acoustic cape
@cosmic nacelle are there any major known failures for this even on GPT-4?
cosmic nacelle
GPT-4 is much better. I’ve seen failures like this, yes. But much less often than 3.5
regal isle
cosmic nacelle
Yes, so most recently I did it like this:
groupchat = autogen.GroupChat(
agents=[user_proxy, assistant, guardrails_agent, distractor_agent],
messages=[],
speaker_selection_method="auto",
allow_repeat_speaker=False,
max_round=12,
)