Wrong. Unless you want to feed the LLM your entire codebase, which is usually infeasible, you need to be able to retrieve relevant context, which relies on understanding the codebase, as Sourcegraph does. Sourcegraph has a product that does precisely this, called Cody.