I noticed Claude code does something similar when "Exploring" the code base, it spawns a subagent using Heroku to get the summary and file locations, which is both faster and it doesn't pollute the context as much.

I wonder how it compares to an vector indexing approach.