I suppose that depends on tooling as well, and the user. Two scenarios:
1. Throw your entire 200k SLOC codebase in and ask for a bug fix.
2. Tokenize your codebase (AST -> SQLite) and have the LLM explore and reason about subgraphs, and then ask for a bug fix.
One takes expertise but yields a better signal to noise ratio. Though, at that level of expertise one encounters what’s effectively the library of babel: you get back what you write and / or already know, and coherent text outlining future events might be wrong until they transpire
1. Throw your entire 200k SLOC codebase in and ask for a bug fix.
2. Tokenize your codebase (AST -> SQLite) and have the LLM explore and reason about subgraphs, and then ask for a bug fix.
One takes expertise but yields a better signal to noise ratio. Though, at that level of expertise one encounters what’s effectively the library of babel: you get back what you write and / or already know, and coherent text outlining future events might be wrong until they transpire