There is a new type of RAG that is powered by LLMs in the loop. The Two that I’ve played with are LightRag and Cognee. As far as I can tell they are mostly identical, but LightRag is more plug and play where Cognee has more setup and more powerful levers to pull. I think LightRag would really pull away if it went all in and automatically synced with a folder on your filesystem. Building these graphs is pretty expensive because you need a smart model to read over every document to extract key features. I’m interested in seeing if Embedding Models can grow much from where they are now and do that extraction themselves, it seems possible based on the latest Chinese Models.