ACTIVE makes dual-vector indexes much cheaper to update—but not uniformly better to search

Dual-vector retrieval has an awkward systems problem: an object may carry two embeddings—image and text, for example—and each query can weight them differently. A new paper, ACTIVE, argues that the harder production problem is not only finding neighbors under those changing weights, but keeping the index useful while objects are inserted and deleted.

The strongest result is on deletion, not search

The evaluation covers five public datasets: OpenImage, Howto100M, CC3M, LAION and Caselaw. Each has 1,000 query vectors except Caselaw, which has 88,300. The default metric is Recall@10; update throughput is measured as operations per second. For deletion tests, the authors remove 1% of vectors in each of 10 batches from a full index. For insertion tests, they begin with a 50% index and add 1% in each of 10 batches. All reported update runs use 32 threads.

Against reinsertion-based repair (RBR), ACTIVE reports 22.64–39.51× higher deletion throughput across the five datasets. Against neighbor-reconnection repair (NRBR), the range is 1.64–2.06×. Its insertion path is less dramatic but still faster: 1.20–1.43× the incremental construction strategy used by DEG. The paper says these gaps remain larger than the standard deviation across three runs.

The mechanism matters. ACTIVE does not apply one repair policy everywhere. It estimates how much connectivity a deletion removes, then chooses among historical-frontier, affected-layer and multi-layer repairs. For insertion and construction, it limits active-range comparisons using a Pareto-angle constraint.

Faster updates do not erase the quality trade-off

The deletion speedup is not free. Against RBR, ACTIVE’s Recall@10 gap ranges from −1.23 to +0.49 percentage points. On Howto100M, RBR is more accurate, although the paper reports RBR is 30.28× slower on deletion there. Against NRBR, ACTIVE improves Recall@10 by 0.21–2.03 points. After insertions, its Recall@10 stays much closer to DEG, with a gap from −0.13 to +0.04 points.

Static search is also mixed. At fixed 0.95 recall, ACTIVE generally beats separate per-attribute indexes as result size grows, but DEG can still be faster on some workloads. Caselaw is a revealing exception: at k=10, the multi-index baseline wins because only about 1,993 coordinate values occur among roughly 7.2 million objects, making the workload behave more like single-vector text retrieval.

What the benchmark establishes—and what it does not

This is evidence that adaptive local repair can sharply reduce update cost for Pareto-frontier indexes while keeping measured recall near stronger repair baselines. It is not yet a universal production verdict. The experiments run on one Ubuntu server with an Intel CPU and 320 GB of memory; the paper does not present an independent reproduction. Its strongest claim is therefore specific: on these five datasets and this batched 10% update protocol, ACTIVE moves the update-throughput frontier, with a small but workload-dependent accuracy cost.