Top-K 并非混合检索的预算
Top-K Is Not a Budget for Hybrid Retrieval
浏览论文内容
中文总结 AI 辅助
针对混合检索中固定截断深度不灵活的问题，提出 DiBud 方法，以访问预算为输入，增量式返回 RRF 排序的精确前缀，在预算内提升认证输出并减少访问次数，实验验证了其高效性。
中文摘要 AI 辅助
现代 RAG 的混合检索通常融合稠密检索器和稀疏检索器的 Top-L 结果，但固定的截断深度可能无法适应变化的查询和语料库。精确融合消除了对固定深度的依赖，但完成指定的 Top-K 仍会产生可变的访问成本。我们提出 DiBud，它直接将访问预算作为输入，并增量式地认证和返回完整列表上 RRF 排序的精确前缀。选择性访问在预算内增加了认证输出，而预算停止则限制了每次请求的访问次数。在五个查询集上的实验揭示了完成精确 Top-20 的长尾成本。在 2048 次访问的预算下，DiBud 在前 100 个位置内的平均认证输出比均衡访问提高了 7.86%。在针对 95% 质量保持进行预算校准后，保留查询保持了平均 nDCG@20 的 95.05%–97.68%，同时比完成精确 Top-20 少使用了 65.92%–99.53% 的访问次数。
英文摘要
Modern hybrid retrieval for RAG typically fuses the Top-$L$ results from dense and sparse retrievers, but a fixed truncation depth may not transfer across changing queries and corpora. Exact fusion removes the dependence on a fixed depth, yet completing a specified Top-$K$ still incurs variable access costs. We present DiBud, which takes an access budget directly as input and incrementally certifies and returns an exact prefix of the RRF ranking over the full lists. Selective access increases certified output within the budget, while budgeted stopping bounds accesses per request. Experiments on five query sets reveal long-tailed costs for completing exact Top-20. At a budget of 2048 accesses, DiBud increases mean certified output within the first 100 positions by 7.86% over balanced access. After budget calibration for 95% quality retention, held-out queries retain 95.05%--97.68% of mean nDCG@20 while using 65.92%--99.53% fewer accesses than completing exact Top-20.
