If this works as advertised, it will make LMM/LLM-RAG systems more effective. I've often found it amusing how as soon as we encounter significant limitations in the architecture, we come running back to search algorithms to help it along. I'm not one of those people that consider this some sort of moral slight against the LLM/LMM's intellectual capacity, my view has always been that it's a system that is or isn't intelligent, not a component. It's just funny considering some of the early "This will replace search!" prognostication which like... yes, it probably will in some sense be an integral part of future search, but the potential is so much more exciting than that.
The worst case performance of this system as described won't be better than the current worst case. It might make the average case more performant, but this paper claims the current solutions don't index at all. That's quite a statement. I would take this with a full cup of salt.