I’m considering something like the following: • chunk by block of some sort - maybe split by heading? paragraph? • create an embedding for each block, store embeddings with ref to which doc it came from • create embedding on search query, then vector search over chunks is this overcomplicating it?