How to Efficiently Explore Noisy Historical Data? Leveraging Corpus Pre-Targeting to Enhance Graph-based RAG