I would speculate that for learning unknown codebases, fine-tuning might work better than relying on context window size.