WARP: An Efficient Engine for Multi-Vector Retrieval

---

title: "WARP: An Efficient Engine for Multi-Vector Retrieval" slug: "warp-an-efficient-engine-for-multi-vector-retrieval" brand: "arXiv" brand_logo: "https://upload.wikimedia.org/wikipedia/commons/9/95/ArXiv_logo_2022.svg" source_url: "http://arxiv.org/abs/2501.17788" category: "curated-resource-lists" featured: false images: []

---

Overview

WARP is a research engine for efficient multi-vector retrieval, targeting systems that store and search multiple embeddings per document, such as modern vector databases used for RAG (retrieval-augmented generation) and semantic search.

• Type: Research paper / retrieval engine
• Domain: Information Retrieval (cs.IR)
• Publication venue: Accepted at SIGIR 2025
• Identifier: arXiv:2501.17788 (cs.IR)
• DOI: 10.48550/arXiv.2501.17788

Features

• Multi-vector retrieval focus

  • Designed for settings where each document is represented by multiple embeddings (e.g., passages, chunks, or token-level vectors).
  • Targets workloads typical in RAG pipelines and semantic search systems.

• Efficiency-oriented engine

  • Aims to improve retrieval performance (e.g., latency, scalability) when handling multi-vector document representations.
  • Tailored to be integrated with or compared against modern vector databases.

• Research-grade implementation

  • Presented as a research engine suitable for experimentation and benchmarking in academic and industrial IR contexts.

• Scholarly integration

  • Indexed and citable via arXiv with a persistent DOI.
  • Linked to a related DOI (10.1145/3726302.3729904), enabling connection to associated SIGIR 2025 resources.

Citation

• arXiv ID: arXiv:2501.17788 [cs.IR]
• Version referenced: v3 (as indicated on the arXiv page)
• DOI: https://doi.org/10.48550/arXiv.2501.17788

Pricing

• Not applicable. This is a research paper and engine published via arXiv; access is free.