Filter and compress context before LLM calls to cut tokens while keeping key information.
Retrieve only the relevant code for a codebase question, within a token budget.
