to answer your question, you'd use something like the knowledge graph to tease out relevant product queries from the conversation and train the model to respond

to answer your question, you'd use something like the knowledge graph to tease out relevant product queries from the conversation and train the model to respond with both the general response but also what you might called the custom intent, which would vary per user. this is what search already does, but making this scale with a language model is probably expensive, hence why google doesn't already release meena