Fidelity synopses cut vector-search cost to 0.72×—by spending a declared miss budget

The number

A new evaluation of approximate vector search reports that a fidelity synopsis reduced median query cost to 0.72× a static plan across all 16 combinations of four embedding models and four query directions. That headline includes the synopsis lookup and comes from a held-out workload—not the sample used to build the selector.[^1]

The test is unusually large and explicit. The authors indexed 5,802,772 LAION image-caption pairs, built synopses on a separate 682,444-pair construction sample, and reported results on 341,071 held-out pairs. Each model was tested in image-to-image, text-to-text, image-to-text and text-to-image search. The plan catalog contained nine configurations spanning HNSW, IVFPQ and OPQ-IVFPQ.[^1]

What the selector actually promises

The result is not “28% cheaper at identical per-query recall.” The default contract sets a fidelity target of 0.7—recovering at least 70 of the exact top 100 neighbors—and a reliability level of 0.9, allowing at most 10% of queries to miss that target. The static baseline is the cheapest single configuration meeting the same workload requirement. The synopsis predicts, before execution, which configuration is likely to suffice for each query and spends unused reliability headroom on cheaper plans.[^1]

On the held-out workload, the budget rule stayed within the 10% miss requirement in every setting. But uncertainty matters: in eight of 16 settings, the upper end of its 95% confidence interval crossed 10%, by at most a little over one percentage point. At the static plan’s lower realized miss rate, the method was cheaper in 12 of 14 reachable settings, with savings of at least 8% in ten.[^1]

Why the result is useful—and bounded

The paper’s strongest design choice is split hygiene. Static-plan selection, synopsis construction and budget tuning happen on the construction sample; the reported cost and fidelity outcomes come from the held-out workload. Costs were measured on dedicated AMD EPYC 7763 nodes with indexes in memory, using 4,096-query batches over 64 threads for headline results.[^1]

The boundary is equally important. This is one LAION workload on one hardware profile, with four embedding models and a hand-built nine-configuration catalog. About half of fidelity variance remains inside synopsis cells, limiting which hard queries the method can identify. The paper also says exact single-query landmark lookup is too slow for 2,048-dimensional models; its 21–32 microsecond lookup headline is batched.[^1]

The finding is therefore narrower than “adaptive vector search is solved.” It shows that a database-style synopsis can turn a declared workload-level error budget into lower ANN cost on a large held-out test. Production adopters still need to decide whether that miss budget is acceptable—and rerun the contract under their own embeddings, query mix, hardware and latency regime.

[^1]: Fidelity Synopses: Pre-Execution Plan Selection for Approximate Vector Search, arXiv v1, submitted Oct. 30, 2026