Research map / Learning theory and algorithms / Combinatorial and search optimization
DESSERT: An Efficient Algorithm for Vector Set Search with Vector Set Queries
Joshua Engels, Benjamin Coleman, Vihan Lakshman, Anshumali Shrivastava
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DESSERT instead of PLAID for k=10 retrieval | ColBERT on MS MARCO development · MS MARCO development, k=10 | latency: 45.1 -> 9.5 | Improved |
| Use DESSERT instead of PLAID for k=10 retrieval | ColBERT on MS MARCO development · MS MARCO development, k=10 | MRR@10: 39.2 ± 1.15 -> 35.7 ± 1.14 | Hurt |
| Use DESSERT instead of PLAID for k=10 retrieval at the higher-latency point | ColBERT on MS MARCO development · MS MARCO development, k=10 | MRR@10: 39.2 ± 1.15 -> 37.2 ± 1.14 | Hurt |
| Use DESSERT instead of PLAID for k=10 retrieval at the higher-latency point | ColBERT on MS MARCO development · MS MARCO development, k=10 | latency: 45.1 -> 15.5 | Improved |
| Use DESSERT instead of PLAID for k=1000 retrieval | ColBERT on MS MARCO development · MS MARCO development, k=1000 | latency: 100 -> 22.7 | Improved |
| Use DESSERT instead of PLAID for k=1000 retrieval | ColBERT on MS MARCO development · MS MARCO development, k=1000 | R@1000: 97.5 ± 0.36 -> 95.1 ± 0.49 | Hurt |
| Use DESSERT instead of PLAID for k=1000 retrieval at the higher-latency point | ColBERT on MS MARCO development · MS MARCO development, k=1000 | latency: 100 -> 32.3 | Improved |
| Use DESSERT instead of PLAID for k=1000 retrieval at the higher-latency point | ColBERT on MS MARCO development · MS MARCO development, k=1000 | R@1000: 97.5 ± 0.36 -> 96.0 ± 0.45 | Hurt |
| Use DESSERT rather than optimized PyTorch brute force | Vector set search on synthetic GloVe sets · Synthetic GloVe sets, N=1000 | speedup: 10-50x | Improved |
| Use DESSERT rather than PLAID | ColBERT passage retrieval · LoTTE, Lifestyle, Technology, Recreation, and Science splits | latency: 2-5x | Improved |
| Use single-vector ScaNN search instead of vector-set retrieval | Passage retrieval · Passage retrieval, corpus unspecified | Recall@1000: 0.77 | Hurt |
Closest papers by meaning
- Searching for and Avoiding Hidden Sets Using Queries with Local Feedback (AAAI 2025 · similarity 0.82)
- MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encoding (NeurIPS 2024 · similarity 0.81)
- LotusFilter: Fast Diverse Nearest Neighbor Search via a Learned Cutoff Table (CVPR 2025 · similarity 0.81)
- Approximate Nearest Neighbor Search with Window Filters (ICML 2024 · similarity 0.81)
- Distribution-Driven Dense Retrieval: Modeling Many-to-One Query-Document Relationship (AAAI 2025 · similarity 0.80)
- Sort Before You Prune: Improved Worst-Case Guarantees of the DiskANN Family of Graphs (ICML 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
