Comment by daemonologist

It's described in the remainder of the point - they use an LLM to generate additional search queries, either rephrasings of the user's query or bringing additional context from the chat history.