Tracert-Retrieval-Augmented Generation: Boosting Multi-Hop Retrieval-Augmented Generation with Direction-Aware Graph Traversal

Tracert-retrieval-augmented generation (RAG) is a novel retrieval-augmented framework designed for efficient, document-level multi-hop reasoning. Unlike conventional RAG systems that retrieve top-k text segments based solely on dense similarity, Tracert-RAG predicts a semantic goal vector from the user query, constructs a local semantic graph from the document embeddings, and employs a direction-aware greedy traversal to identify reasoning paths toward the goal. This system eliminates the inflexibility of symbolic graph traversal and the inefficiency of manual query reformulation. On literary analysis tasks from Pride and Prejudice, Tracert-RAG outperforms standard RAG and graph RAG baselines in answer quality, inference speed, and interpretability. Specifically, it achieves the highest average answer quality (8.05 out of 10) while reducing indexing time by a factor of 80 compared to graph RAG methods.