GRACE: Efficient Group-Aware Refinement for Approximate Certified Embedding Queries
Year: 2026
Type: article
Abstract: Modern analytical applications increasingly manage objects that combine structured attributes with unstructured content represented by learned embeddings. Although vector indexes support efficient semantic retrieval, many analytical workloads require grouped aggregates over embedding-defined neighborhoods together with structured predicates. A conventional retrieve-then-aggregate plan only applies predicates and computes aggregates. This separation can incur substantial vector scoring and payload-access costs, while retrieval recall alone provides no deterministic guarantee on downstream grouped aggregates. We study certified grouped aggregation over threshold- and top-k-defined semantic neighborhoods and present GRACE, a group-aware framework that integrates semantic retrieval, structured filtering, and aggregation. We prove that the returned count, sum, and average intervals contain the exact grouped aggregates under conservative synopsis and numerical assumptions. Experiments on million-scale audio, video, and text embedding workloads show that GRACE improves query throughput over retrieve-then-aggregate baselines, achieves exact top-kmembership after normal completion, preserves full empirical certificate coverage, and incurs moderate preprocessing and memory overhead.
Source: IEEE Access
Authors: Yanting Sun, Yuntao Sun, Xuehua Zhang
Institution: Shandong College of Electronic Technology
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7225660669