Not as big a deal as you thin. Embedding chunking and retrieval already works well for any arbitrary outside knowledge (database, books, Codebase etc)