we instruct model to generate product recommendation directly in response -> so all these are part of context when using history (possibly they're summarized by LLM -> so may not be an issue)