Dynamic-Query Robustness of ANN Indexes under Time-Indexed Drift
Document Type
Presentation
Publication Date
Spring 4-24-2026
Abstract
Approximate nearest-neighbor search is a central retrieval primitive in dense question-answering and retrieval-augmented generation systems. Existing ANN evaluation protocols typically measure recall, latency, throughput, and search-effort sensitivity under a fixed-query assumption: a query vector is submitted to an index, approximate neighbors are retrieved, and the result is compared with exact nearest-neighbour ground truth. This assumption is appropriate for conventional vector-search benchmarking, but it is less complete for multi-step, distributed, and agent-controlled retrieval pipelines in which the retrieval-facing query may be refined, recomputed, or displaced across execution steps. This paper introduces a time-driven dynamic query evaluation framework for ANN search. The framework models query-state movement as stochastic corruption on the unit hypersphere, where elapsed compute time or network round-trip time determines the drift budget. The evaluation is instantiated over HNSW and IVF-Flat using MS MARCO passage embeddings encoded with BAAI/bge-base-en-v1.5. Results show that fixed geometric drift reduces recall without changing the preferred operating point, whereas compute-time and network-time coupled dynamic queries shift the best ANN settings towards shallower search. Matched no-drift controls confirm that the effect arises from coupling elapsed time to query movement rather than from repeated retrieval alone. The findings indicate that latency in ANN-indexed QA/RAG systems should be treated not only as a performance cost but also as a retrieval-quality variable when query state is dynamic.
Program or Discipline Name
Computer Information Sciences
Secondary Program or Discipline Name
Computer and Information Sciences
Recommended Citation