BUCKETS: A Role-Aware Semantic Graph with Hierarchical Aggregation for Efficient Document Retrieval

We present BUCKETS, a document retrieval system combining semantic graphs with role-based hierarchical aggregation for question-answering tasks. BUCKETS organizes chunks into semantically coherent buckets with domain-specific roles (TABLE, RESULT, METHOD, DEFINITION), using centroidbased representations for efficient similarity computation. On technical documents, BUCKETS achieves 9.3× speedup over FAISS and 3× over semantic graphs while maintaining 100% precision (vs. 67-83% baselines). Role-aware retrieval reduces LLM inference time by 73% through superior context quality, suggesting that semantic aggregation with structural awareness offers a promising direction for specialized domain retrieval.