How to keep a RAG knowledge base fresh without re-embedding everything

Most RAG pipelines I look at are refreshed by brute force: re-crawl the docs site every night, re-chunk everything, re-embed everything, overwrite the index. It works, eventually, but it's wasteful an