pgvector's cost-based planner picks approximate scans where exact scans hit perfect recall

pgvector's result sits in the cost model. The optimizer prices an approximate index scan against an exact sequential scan and, the authors report, frequently picks the approximate one even where the sequential scan would have returned every true neighbour at comparable latency ^[6]. Recall is spent and no latency comes back ^[15]. The paper's remedy is not a new index type: it synthesises guidelines for selecting index types and for configuring query optimizers on hybrid search workloads ^[10].

I would expect the Milvus finding to be the same effect from the other side. An engine that can fall back to exact evaluation over the rows that survive a predicate has a floor under its recall, and the abstract attributes Milvus's recall stability to exactly that hybrid approximate/exact execution ^[5].

Before carrying either result into your own cluster, look at what the filters correlate with. The authors propose Global-Local Selectivity, a correlation metric for the relationship between filters and query vectors ^[4]. That relationship decides whether a predicate thins the corpus roughly uniformly or removes the region the query vector points at. Their production examples are ordinary ones: a medical RAG query for similar symptoms restricted to an age range, a legal search restricted to a jurisdiction ^[18]. IVFFlat beating HNSW on low-selectivity queries ^[7] transfers if your predicates sit on your embedding space the way MoReVec's do, and the paper hands you the metric to check that instead of asking you to take the ordering on faith.

There is also the definitional problem the authors name in their own motivation. They write that the literature lacks a unified taxonomy for these strategies, which has produced ambiguous definitions and inconsistent implementations across systems ^[13]. Two engines can both advertise pre-filtering and do different work behind the word. That matters for anyone comparing vendors on a slide, because the generic strategies are what schema-agnostic engines are left with ^[12].

The supplied abstract and introduction do not include per-engine latency or recall figures, and do not say whether low selectivity means few rows surviving the filter or many. The authors extended ANN-Benchmarks to cover filtered vector search and released it ^[9], so the pgvector plan finding is checkable against another dataset. Until someone does that, it is one group's measurement on the two-table dataset they built for it ^[3].