This is usually a result of the context sent to the LLM being too small. Could you increase the

max_bytes

setting in the conversation model to as high as say 16K?