Internal Conversation Formatting Messes Up Conversation Sequence Turns and Gives Odd Responses

KawaneRio

The internal conversation sequence seems to be messed up. Model ignores the last asked question and goes back to answer something that has been asked previously. Upon looking into the "thinking", it is revealed that the "current prompt" is incorrectly relayed to the model.

It is heavily implied that my prompt is internally "formatted oddly" and that my input appears in the "Information provided internally by ResearchAgent" section when that shouldn't be the case.

Make sure my prompts are relayed to the model clearly with proper formatting internally so that the model does not get confused what the user's last input was.

Carolly

My experience with this is that it is caused by the Assistant failing to differentiate between current-turn input and previous-turn context provided by the agent responsible for this (which I have nicknamed ContextAgent for the sake of argument) - in longer conversations it provides a block of text labeled chat_history.md but the Assistant can't always tell the difference between the content of that file and the active conversation (it also sometimes can't tell "who's who" in the conversation.)