vector similarity search

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.

Existing embedding-based retrieval systems employ fixed recall budgets, leading to insufficient recall for head queries and degraded precision for tail queries—rooted in the frequentist paradigm of loss functions. This paper proposes a query-aware probabilistic retrieval framework that abandons fixed truncation and instead models the query-specific distribution of candidate cosine similarities. By learning a query-conditioned cumulative distribution function (CDF), the framework dynamically determines similarity thresholds per query. For the first time, it jointly optimizes precision and recall for both head and tail queries within a unified probabilistic framework. Experiments across multiple industrial retrieval benchmarks demonstrate significant improvements in overall effectiveness. Ablation studies validate the efficacy of explicitly modeling head–tail disparities. The core innovation lies in elevating threshold selection from heuristic, static configuration to a principled, query-adaptive, probabilistic decision process.

Traditional vector retrieval relies on approximate nearest neighbor (ANN) search, which often yields semantically redundant results and fails to meet the diversity and contextual richness requirements of applications such as retrieval-augmented generation (RAG) and multi-hop question answering. To address this, we propose a novel paradigm—“Semantic Compression and Graph-Enhanced Retrieval”—that introduces submodular optimization into vector retrieval for the first time. We formalize semantic compression to maximize information coverage while explicitly suppressing redundancy. Leveraging information-geometric similarity metrics and k-nearest neighbor (kNN) graphs, we construct a multi-hop semantic search framework, further augmented with knowledge graph integration for structured semantic querying. Our method supports hybrid indexing and significantly improves semantic diversity and coverage in high-dimensional embedding spaces, outperforming state-of-the-art ANN baselines. The implementation is open-sourced, advancing research toward semantics-centric vector search.

This work proposes a novel approach to large-scale retrieval that circumvents the prohibitive cost of full reranking by constructing query and item embeddings derived from the outputs of a reranker. Specifically, it leverages relevance scores assigned by a heavyweight reranker over a set of support items to generate lightweight embeddings, thereby enabling the reranking model to directly guide embedding learning—a capability demonstrated here for the first time. Under mild conditions, the method is theoretically shown to approximate arbitrarily complex similarity functions. Through systematic investigation of support item selection strategies and integration with approximate nearest neighbor search, the approach significantly improves candidate set quality across multiple academic and industrial datasets while maintaining computational efficiency.

This paper identifies a fundamental theoretical limitation of the vector embedding retrieval paradigm: even for simple queries, the embedding dimension strictly bounds the number of distinguishable document subsets—a bottleneck intrinsic to the representation and unresolvable via larger models or improved training data. Method: Drawing on statistical learning theory, the authors formally derive an upper bound on the expressive capacity of single-vector embeddings, then construct LIMIT, the first benchmark explicitly designed to probe this theoretical limit; they further propose a minimal empirical framework with k=2 and fully parameterized embeddings for boundary testing. Results: Experiments show that state-of-the-art embedding models underperform significantly relative to the derived theoretical ceiling on LIMIT, directly challenging the prevailing assumption that scaling model size alone suffices to overcome retrieval limitations—and thereby establishing a structural capacity ceiling for the single-vector paradigm.

In vector retrieval, inconsistent embedding quality causes significant query-level performance fluctuations, while existing methods lack lightweight, training-free capabilities for per-query performance prediction. To address this, we propose Q-Robust—a fine-tuning-free, lightweight framework that jointly models geometric robustness (quantifying local stability) and neighborhood density (capturing semantic compactness) in the embedding space to accurately predict retrieval effectiveness for individual queries. Our analysis reveals systematic patterns in query-specific embedding quality distributions, enabling dynamic, adaptive retrieval strategies. Evaluated on four standard benchmarks, Q-Robust achieves an average 9.4±1.2% improvement in Recall@10 over strong baselines, with prediction overhead accounting for less than 5% of retrieval latency—demonstrating superior accuracy, efficiency, and practicality.

Existing generative retrieval methods in recommendation and advertising systems are hindered by quantization dependencies, unstable identifier vocabularies, and insufficient query-item alignment. This work proposes the Embedding-based Generative Retrieval (EGR) framework, which unifies generative retrieval and end-to-end embedding learning within a single shared large language model for the first time. EGR directly learns dense embeddings from item metadata and user interaction histories, enabling query-item alignment in a unified vector space. By eliminating reliance on traditional semantic IDs, the approach supports cold-start scenarios and multimodal inputs while simplifying system architecture. Experiments demonstrate that EGR outperforms current baselines on the Amazon Reviews dataset and exhibits strong scalability and cold-start performance on Snap DPA data, yielding a 2.91% improvement in online conversion rate upon deployment.

This study investigates the impact of embedding dimensionality on performance in dense retrieval and its limitations as task complexity increases. Through systematic experiments across models of varying scales, the work presents the first empirical evidence that retrieval performance follows a power-law relationship with embedding dimensionality. Building on this observation, the authors propose predictable scaling laws based solely on dimensionality or jointly on model size. Using dense retrieval architectures, approximate nearest neighbor search, and large-scale comparative evaluations, they demonstrate that in task-aligned scenarios, performance improves with higher dimensionality—albeit with diminishing returns—whereas in misaligned tasks, excessive dimensions degrade performance. These findings offer both theoretical grounding and practical guidance for selecting optimal embedding dimensions in efficient retrieval systems.

Generative retrieval (GR) excels in supervised settings but suffers from poor generalization in zero-shot information retrieval (IR). This paper introduces ZeroGR—the first general-purpose generative framework for zero-shot retrieval over heterogeneous documents (e.g., text, tables, code). To address the challenge of format-agnostic semantic alignment, ZeroGR employs: (1) an instruction-tuned query generator that unifies modeling across diverse document modalities; and (2) a reverse-annealing decoding strategy to enhance generation stability and relevance. The framework is end-to-end optimized and seamlessly integrates large language models. Evaluated on the BEIR and MAIR benchmarks under strict zero-shot settings, ZeroGR significantly outperforms both dense retrievers and existing generative baselines, achieving state-of-the-art performance. Results demonstrate ZeroGR’s strong cross-task generalizability, scalability to unseen tasks, and robustness across document formats—without any task-specific fine-tuning.

Zero-shot dense retrieval suffers from semantic ambiguity among top-ranked documents due to the absence of relevance labels for training queries. To address this, we propose a training-free representation sharpening framework that enhances document embeddings via context-aware representation refinement during indexing—improving semantic discriminability without modifying the underlying retriever. Our method is agnostic to any pre-trained dense retriever and incorporates an approximation strategy to balance effectiveness and computational overhead. Evaluated across 20+ multilingual zero-shot benchmarks—including the BRIGHT benchmark—our approach achieves new state-of-the-art performance. Its approximate variant retains over 90% of the full method’s gains while incurring zero additional inference cost. The core contribution is the first unsupervised, fine-tuning-free, and computationally efficient document representation sharpening technique, which significantly alleviates semantic confusion in zero-shot dense retrieval.

This work addresses a critical limitation in traditional approximate nearest neighbor (ANN) retrieval evaluation, which relies solely on recall and fails to distinguish semantically relevant from irrelevant neighbors, often misrepresenting retrieval quality. To overcome this, the authors introduce Semantic Recall—a novel metric that quantifies only those semantically relevant results theoretically recoverable via exact search—and propose Tolerant Recall as an efficient proxy for practical evaluation. This is the first systematic integration of semantic relevance into vector retrieval assessment, revealing the pervasive sparsity of relevant results within embedding spaces. Experimental results demonstrate that the proposed metrics more accurately reflect real-world retrieval effectiveness, and algorithms optimized under this framework achieve superior trade-offs between cost and quality.