LLM Context Is Essentially a Search Problem

It wasn't long ago that RAG and vector databases were seen as the definitive answer for LLM context. The thinking was architectural, a sort of external brain bolted onto the model.

There’s a growing sense that the most effective methods are much simpler: give the model a directory of plain text files and let it search (e.g., Andrej Karpathy's LLM+Obsidian knowledge base). It reframes the whole issue. The bottleneck may not have been the model’s ability to reason, but simply our ability to get it the right information in the first place. We were building elaborate memory palaces when what it really needed was a library card and a fast search index.