RAG is Dead. Again. (Claude Agent SDK + Memory)

Agentic File Search is a very interesting idea for RAG, and from my own experience using it, I found the retrieval accuracy to be quite impressive. However, I think the biggest challenges are latency and token usage. In domains where the volume of documents is massive and documents are continuously updated or removed, operating this kind of pipeline in a stable and scalable way is not easy. I believe Graph RAG has similar limitations as well. In environments where documents exist at large scale and change continuously, the cost of maintaining and synchronizing the graph becomes significant, which makes it difficult to operate reliably at a production level. Personally, I’m curious whether you have any ideas for reducing latency in these kinds of systems.

GraphRAG already solves the issues with classic RAG.

Nice. Would be cool to choose the agent.

If it’s already read the files, why does it need semantic search?

Will Anthropic block this as well soon?

is it open source and free and uncensored, why is it not integrated in a jarvis like linux distro, download and install onto a usb key or external usb drive, boot from it, checks the hw and moves in with audio in and out by default, permanent memory rag and ability to work with webcams and do tasks in browser...from than on it learns on its own

🥱

IMHO, your chuncking part is too simple and could be way more efficient.

First "RAG is dead" video in 2026 👏