MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval

MCompassRAG is a new retrieval framework designed to enhance retrieval-augmented generation (RAG) systems by integrating topic metadata into the retrieval process. It employs a lightweight retriever trained via LLM-teacher distillation, allowing for topic-aware retrieval without additional LLM calls, resulting in an average information efficiency improvement of 8.24% and over five times lower latency compared to existing RAG baselines across six complex benchmarks. This approach is significant for practitioners as it balances precision and efficiency in document retrieval for deep research tasks, potentially reducing operational costs and enhancing the quality of retrieved evidence.