pinecone

Cloud-native transactional databases (e.g., Azure Cosmos DB) lack efficient vector search capabilities. Method: This work pioneers the native integration of DiskANN—a high-performance disk-based indexing library—into a NoSQL engine, enabling unified storage and strong consistency between vector indexes and primary data. We propose a distributed index synchronization protocol, adaptive sharding, and memory-mapped file optimizations. Contribution/Results: On 10M vectors, the system achieves <20 ms P95 latency, high recall, and stable update throughput. Compared to Zilliz and Pinecone Serverless, query cost is reduced by 15× and 41×, respectively, while supporting elastic scaling to 1B–10B vectors. Crucially, it retains cloud database advantages—including high availability, durability, and horizontal scalability—without requiring a dedicated vector database, establishing a new production-grade paradigm for semantic search that is low-cost, low-latency, and strongly consistent.

This study addresses the lack of systematic evaluation of hybrid search mechanisms that combine semantic retrieval with metadata filtering in existing vector databases. We propose a novel relevance metric, Global-Local Selectivity (GLS), construct MoReVec—the first benchmark dataset supporting filtered retrieval—and extend ANN-Benchmarks to enable unified evaluation of hybrid search performance. Through comprehensive experiments integrating diverse filtering strategies into FAISS, Milvus, and pgvector with IVFFlat and HNSW indexes, we demonstrate that engine-level algorithmic integration critically governs performance: Milvus achieves more stable recall via hybrid execution, pgvector’s optimizer often selects suboptimal query plans, and IVFFlat outperforms HNSW under low-selectivity queries. Our findings culminate in practical configuration guidelines that offer both theoretical insights and actionable recommendations for efficient hybrid search deployment.

This paper addresses performance bottlenecks of cloud-native vector search over remote storage by systematically comparing cluster-based and graph-based indexes under high-concurrency, high-recall, and high-dimensional workloads. Through a unified benchmarking framework, we reveal the graph index’s substantial advantages in latency-sensitive and cache-constrained scenarios, and identify the root cause of local tuning parameter failure in cloud environments: a fundamental mismatch between index access granularity and cloud storage caching mechanisms. To resolve this, we propose a cloud-aware graph index parameter redesign method and an index-cache co-optimization strategy that dynamically adapts query granularity and data-fetch patterns to available cache capacity. Experimental evaluation on representative cloud configurations demonstrates an average 37% reduction in query latency and a 5.2% improvement in recall. This work is the first to rigorously characterize the cloud-native applicability boundaries of these two dominant indexing paradigms.

Existing graph-based indexes (e.g., HNSW) suffer from poor connectivity, low recall, and high latency for nearest-neighbor search under complex hard filtering predicates. This paper proposes a multi-index ensemble architecture coupled with a learnable three-dimensional analytical model to enable dynamic, query-time selection of the optimal index. Our approach integrates workload-aware index packaging, parallel multi-index construction, and a lightweight HNSW variant—achieving high recall while significantly improving efficiency. Experiments show that our method accelerates queries by up to 8.06× over state-of-the-art baselines, reduces indexing time to just 1% of conventional methods, and incurs only 2.15× the memory overhead of standard HNSW. The core innovation lies in decoupling filtering logic into a collaborative multi-index mechanism and employing a learnable analytical model to drive real-time index selection—thereby overcoming the performance bottlenecks inherent to constrained graph traversal in predicate-aware ANN search.

To address the growing demand for efficient billion-scale embedding vector retrieval driven by AI applications, this paper systematically analyzes Faiss’s trade-offs among accuracy, latency, and memory overhead, and proposes a modular index architecture with hardware-aware optimizations. Methodologically, it introduces a unified indexing framework integrating product quantization (PQ), inverted file (IVF), and multi-level indexing; a standardized API supporting CPU/GPU heterogeneous backends; and low-level optimizations including SIMD acceleration, memory-mapped I/O, and optimized quantization encoding. Evaluated on standard benchmarks, the approach achieves millisecond-latency similarity search over billion-vector datasets, delivering 2–5× higher throughput than prior systems. The resulting infrastructure has been deployed in production recommendation, retrieval, and multimodal search systems at Meta, Netflix, and other industry leaders, establishing a scalable, high-performance paradigm for large-scale vector databases.

Existing VSS benchmarks overemphasize the recall–latency trade-off and rely on a single distance metric, neglecting the actual impact of retrieval results on downstream tasks—leading to evaluation distortion. This paper introduces Iceberg, the first end-to-end, task-effect-driven VSS benchmark suite, which uncovers three fundamental performance bottlenecks: embedding loss, metric misuse, and data distribution sensitivity. We propose a task-centric evaluation paradigm, an information-loss funnel model, and a comprehensive benchmark comprising eight annotated datasets ranging from 10M to 100M samples. Our multidimensional, task-driven evaluation framework integrates cross-modal distribution modeling, application-aware metric re-ranking, and meta-feature-guided decision tree learning. Re-evaluating 13 state-of-the-art methods across eight real-world scenarios reveals significant misalignment between conventional recall-based rankings and actual task performance. Iceberg delivers interpretable, workload-adapted guidance for VSS method selection.

This work addresses the inefficiency of existing vector join methods in threshold queries, which suffer from redundant index traversals and excessive distance computations. To overcome these limitations, the authors propose a unified framework that integrates three key innovations: a soft work-sharing mechanism, a merged graph index that jointly embeds query and data vectors, and an adaptive hybrid search strategy tailored for out-of-distribution queries. Evaluated on eight benchmark datasets, the proposed approach consistently outperforms state-of-the-art methods, achieving substantially lower computational overhead while maintaining high recall. The results demonstrate a superior trade-off between efficiency and accuracy for approximate threshold vector joins.

This work addresses the limitation of existing vector databases, which support only categorical or numerical attribute filtering and struggle to efficiently handle joint queries involving sequence pattern constraints—such as substring matching—and approximate nearest neighbor (ANN) search. To bridge this gap, we propose VectorMaton, a novel index structure based on augmented suffix automata that enables deep integration of sequence pattern predicates (e.g., LIKE/CONTAINS) with ANN search for the first time. VectorMaton achieves significant query efficiency gains while maintaining an index size close to that of the raw data. Experimental results on real-world datasets demonstrate that, at comparable accuracy, VectorMaton improves query throughput by up to 10× over baseline methods and reduces index size by up to 18×.

This work addresses the operational complexity and architectural coupling introduced by conventional vector similarity search systems in compute-storage disaggregated environments, which typically rely on a separate indexing layer. To resolve this, the authors propose deeply integrating a distributed approximate nearest neighbor (ANN) index into the Apache Iceberg table format. The approach leverages Puffin sidecar files to store sharded Vamana graphs and employs Iceberg’s snapshot mechanism to atomically bind indexes with data, enabling versioning and time travel. A coordinator-executor protocol is introduced, featuring a hierarchical probing strategy where coordinators cache compact centroid indexes while executors store large graphs on SSDs. The design reuses Iceberg’s REST catalog optimistic concurrency control for index commits. This solution is the first to host billion-scale vector graph indexes within Puffin, achieving a favorable trade-off between recall and latency while preserving Iceberg’s native transactional capabilities and significantly reducing system complexity.

This study systematically evaluates the practical efficacy and applicability boundaries of distance comparison operations (DCOs) in vector similarity search, addressing the critical question of whether DCOs are suitable for production-grade vector databases. Through the first comprehensive benchmark of eight DCO algorithms across diverse hardware platforms—including CPU (with and without SIMD) and GPU—on ten billion-scale datasets with dimensions up to 12,288, the work reveals that DCO performance is highly sensitive to data dimensionality and hardware configuration. Notably, under out-of-distribution queries, certain DCOs even underperform full-dimensional scanning, indicating they currently fall short of production deployment requirements. Nevertheless, the study identifies promising potential for DCOs to accelerate index construction and data updates.