Right. You feed the text chunks (from the matched embeddings) to a generative LLM to do the extractive/summarization part.