The idea that more context is good is highly ...contextual... and that seems lost on these tools. It isn't just a hallucination, it's that the system isn't designed to constrain the model in order to achieve the desired outcome. I'm left to prompt engineer the constraints, e.g. "Do not use content from documents A, B, C. Only use content from X, Y, Z," which is too much hassle for the value being provided by the generation.