Verified, then logged.
Sun 22 Nov 2026 · 00:34 KST / 15:34 UTC한국어 | ENSearch /

01 / AI · Claim vs. Check · Analysis

MCGI adjusts vector search effort to data geometry

MCGI uses distances calculated during search to adjust search effort, with performance measured against DiskANN on the same index.

Five open laptops sit in a row on a long table in a Brussels office, only two of their screens glowing brightly, as an out-of-focus reviewer seen from behind looks over them. AI-generated image

The researchers behind MCGI reported search throughput of up to 1.43 times that of DiskANN, with both methods using the same index at matched high recall—the proportion of correct results retrieved—in researcher-published benchmark figures.

According to the paper, MCGI varies the search effort allocated to each query. It uses distances already calculated during the search to make that adjustment, without additional disk reads for the estimate.

The paper’s comparison holds the index and recall conditions constant. Those conditions matter when assessing how much throughput changes when search effort is allocated differently.

Search effort follows data geometry

The paper identifies a mismatch between straight-line distances and distances along the underlying shape of the data. A search that repeatedly follows nearby candidates can stray from that structure.

The researchers used local intrinsic dimensionality, or LID, a measure of the complexity of nearby data, to allocate search effort. MCGI gives queries in more complex regions longer candidate lists and lets those in simpler regions terminate earlier.

Search throughput on the same index

According to the paper, the baseline is DiskANN using the same standard Vamana index, the structure used to organise vectors for search. MCGI estimates LID around each query from distances calculated during its own search and adjusts that query’s candidate-list size once.

The researchers also compared performance at matched high recall on the billion-scale SIFT1B and Deep1B datasets, reporting throughput of up to 1.24 times that of DiskANN in researcher-published benchmark figures. The paper’s comparison therefore concerns how search effort is distributed within the same index.