Comment by visarga

Wondering if anyone considered the idea of saving search stubs and using local LLM as agents with it. Could cut the search engine completely from the loop.

The idea is simple: collect search stubs, short documents with routing information about a topic. When you want to search something your agent finds the right stub and looks up where on the web to go find that information. Your information is always fresh, while a search stub changes slowly, over years, because the entry points tend to be more stable.

If we use a 4B local model and another 4GB for search stubs it could be portable enough to download and install. As you use it you also generate your own search stubs on top of the generic package. A stub could contain links to news feeds, high quality hubs, search engines and of course, actual websites. High quality search stubs can be generated with any frontier LLM piggy backing on its agentic search capabilities.

I think the idea of managing a collection of search stubs as a replacement for centralized search engines is important because it would wean us of one of the last centralized points of the internet. Google plays many ranking games on top of users and publishers, serving their own interests first. I want out of that arrangement.

local LLM + local stub index + proxy == anonymity