[name] asked: How did you manage to reduce the context size fed into the LLM by about a quarter?