isn’t this just semantic chunking? There’s been white papers and already implementations in Langchain and llama index you could look through.