I think the problem with vector search is that it’s unbounded - is the 3rd result relevant at all? What about the 4th?
On top of that, I believe it’s clear with how Cursor and Claude converged on regex search after using vectors that asking LLMs to come up with similar terms to regex is more effective than searching for the user’s original term over a vector db.
