hybrid retrieval

This study investigates whether retrieval fusion techniques—commonly adopted in real-world retrieval-augmented generation (RAG) systems, such as multi-query retrieval and reciprocal rank fusion—consistently improve end-to-end answer quality under practical deployment constraints. Conducted within an enterprise knowledge-base RAG pipeline, the evaluation is performed under fixed retrieval depth, reranking budget, and latency limits. While retrieval fusion enhances initial recall, it fails to translate into improved Top-k accuracy after subsequent reranking and context truncation; notably, Hit@10 declines from 0.51 to 0.48 and incurs additional latency. These findings challenge the prevailing assumption of the default efficacy of recall-oriented fusion strategies, revealing diminishing returns in production settings where downstream processing and system constraints critically shape overall performance.

This study addresses the challenges of financial document retrieval, where queries are often short and contain abbreviations, and relevant evidence is scattered across lengthy documents and tables, limiting the effectiveness of conventional sparse and dense retrieval methods. To overcome these issues, the authors segment documents into passages aligned with the encoder’s input window to fully cover annotated evidence spans, combine BM25 with a compact dense retriever, and apply Reciprocal Rank Fusion (RRF) to enhance performance. The work further reveals that passage length introduces bias in dense retriever evaluation, demonstrates that training-free RRF outperforms uniform-weight fusion, and investigates three lightweight adaptive routing strategies. On the revised FinDER dataset, fixed-weight fusion improves Hit@10 by 28%; however, despite a theoretical improvement margin of 21.8%, the adaptive approaches fail to achieve statistically significant gains.

Existing hybrid search systems lack systematic empirical analysis of trade-offs among lexical and semantic retrieval components—i.e., retrieval paradigms, fusion strategies, and re-ranking methods—leading to complex, suboptimal configurations. Method: We introduce the first benchmark framework tailored for advanced hybrid architectures, conducting systematic evaluation across 11 real-world datasets, covering four retrieval paradigms, their combinations, and re-ranking strategies. Contribution/Results: We identify a “weakest-link” effect in hybrid pipelines and propose a data-driven configuration mapping method. Crucially, we find Tensor-based Re-ranking Fusion (TRF) achieves both high efficiency and strong semantic modeling under low-resource conditions, overcoming traditional fusion bottlenecks. Experiments reveal that hybrid performance is severely constrained by imbalanced path quality; optimal configurations are highly dependent on dataset characteristics and resource constraints. TRF significantly improves the effectiveness–cost trade-off, outperforming state-of-the-art baselines across diverse settings.

To address the challenge of identifying tail-labels in eXtreme Multi-Label Text Classification (XMTC) caused by highly imbalanced (long-tailed) label distributions, this paper proposes a sparse-dense retriever fusion framework. Methodologically, it jointly leverages sparse retrievers (e.g., BM25), which excel at lexical exact matching, and fine-tuned dense retrievers (e.g., BERT), which capture semantic similarity, within a unified embedding space. Candidate labels are retrieved via approximate nearest neighbor search, and a ranking-based fusion strategy adaptively weights outputs from both retrievers. Crucially, the framework exploits their complementary strengths without requiring retraining. Experiments across multiple XMTC benchmark datasets demonstrate that the method consistently outperforms individual retriever baselines: it significantly improves recall for tail-labels while preserving accuracy on head-labels, yielding substantial gains in tail-label F1-score and overall Precision@K.

This work addresses the lack of systematic design principles for neural retrieval systems that balance efficiency and effectiveness. It proposes the first vertically layered four-tier framework—spanning representation, granularity, orchestration, and robustness—to structurally characterize key design decisions at each layer and their interdependencies. By integrating Bi- and Cross-encoder architectures, atomic and hierarchical chunking strategies, multi-stage re-ranking, agent-based decomposition, and domain generalization techniques, the study elucidates the mechanistic impact of each design choice on system performance. This approach effectively mitigates critical challenges such as information bottlenecks, semantic blind spots, and temporal drift, thereby offering a practical and actionable optimization pathway for building efficient and robust embedded retrieval systems.

This work addresses the challenge of effectively fusing multiple heterogeneous retrieval channels under strict latency constraints to optimize business metrics such as user conversion. We propose a channel-aware unified learning-to-rank framework that formulates multi-channel result fusion as a query-dependent multi-objective ranking problem, jointly optimizing for click-through, add-to-cart, and purchase outcomes. The approach explicitly incorporates channel-specific signals and users’ short-term behavioral sequences, and leverages query-adaptive fusion strategies alongside cross-channel interaction modeling to overcome the limitations of conventional fixed-weight fusion methods. Online A/B experiments demonstrate that the system achieves a 2.85% improvement in user conversion rate while maintaining a p95 latency below 50 milliseconds, and has been successfully deployed in the production environment of Target.com.

This work addresses the challenge of effectively fusing heterogeneous scores—such as vector similarity and graph-based relevance measures like personalized PageRank—in graph-augmented retrieval, where distributional mismatches hinder integration. To resolve this, the authors propose a calibration method based on Percentile Rank (PIT) normalization, which maps disparate scores onto a unified, dimensionless scale while preserving magnitude information and enabling stable alignment. Combined with linear and Boltzmann fusion strategies, the approach significantly improves last-hop retrieval performance in multi-hop question answering. On MuSiQue and 2WikiMultiHopQA benchmarks, it achieves LastHop@5 scores of 76.5% and 53.6%, respectively, substantially outperforming existing baselines.

Traditional hybrid retrieval relies on fixed top-L truncation to fuse dense and sparse results, which often introduces ranking bias and struggles to adapt to dynamic changes in queries and corpora. This work proposes Exact Adaptive Hybrid Retrieval (EAHR), the first method to achieve exact adaptive fusion—identical to full-list weighted reciprocal rank fusion (RRF)—without predefining a top-L cutoff. EAHR dynamically adjusts per-channel retrieval depth and continues retrieval only when it impacts the final top-K results. By integrating per-vector scalar quantization and posting block-max techniques, EAHR enables recoverable exact ranking and incorporates a fusion-bound pruning strategy. Experiments demonstrate that EAHR exactly reproduces full top-20 results across five datasets and 150 query–snapshot combinations, achieving 23.35–30.28× speedup over exhaustive batch processing on average.

This work addresses the challenge of ineffective reranking in dense retrieval systems under zero-shot scenarios, where supervised signals are absent. The authors propose DART, a novel method that performs lightweight adaptive training at test time to refine reranking. Specifically, DART generates pseudo-labels from top- and bottom-ranked documents in the initial retrieval results and fine-tunes the bilinear scoring matrix via a small number of gradient updates, guided by a confidence-weighted margin loss and a cross-query momentum buffering mechanism. Requiring no additional annotations, DART achieves an average relative improvement of 2.1% in NDCG@10 across six BEIR benchmarks, with less than 10ms added latency per query.