Search the eligible population without losing its routes

Let eligible IDs be those satisfying a fixed predicate at the query’s selected snapshot. The desired answer is the smallest-distance k items within that set under a defined tie order. If only one item is eligible, a request for ten cannot return ten valid unique documents. Declare how short eligible populations are represented and how recall is scored. Mixing those cases into a denominator of ten without explanation confuses corpus size with approximation failure.

Postfiltering first retrieves an unfiltered candidate set and then applies the predicate. It is simple, but a fixed oversampling factor provides no general completeness guarantee. If every nearby candidate fails the predicate, many farther eligible results may be missed. An exact scan over a very selective eligible subset can be both simpler and faster than exploring a large unfiltered graph, provided eligibility lookup and vector access are available and their costs are counted.

Imagine the only route from an entry point to an eligible document passes through an ineligible point. Removing that intermediate point from traversal disconnects the route, even though allowing traversal while suppressing its output would preserve it. Some filtered graph algorithms account for this distinction; others use specialized structures or fallback search. The supported behavior depends on the actual index and implementation. A generic label filter is not a proof that the induced eligible subgraph is navigable.

Faiss exposes ID selectors through supported per-query search parameters, and their interaction with scanning depends on the index and selector type. hnswlib 0.8.0 exposes a label-based query filter with documented Python performance considerations. Verify these concrete APIs rather than assuming a database-style metadata planner exists inside every ANN library. Keep predicate evaluation, accepted output IDs, and measured candidate effort visible so a performance optimization cannot silently change the eligible population.

Construct ground truth by applying the predicate before exhaustive scoring, using the same document versions and metadata seen by the query. Include predicates with different selectivities and correlations with embedding geometry. A tenant, language, or publication-time filter may cluster strongly in the vector space, so a global selectivity average is a weak predictor of the current candidate stream. Measure completeness and latency separately for these slices, including empty eligible sets and boundary cases.

When the predicate expresses access visibility, validate returned IDs against the authoritative visibility snapshot before exposing document contents. This output condition does not imply that every internal graph node must be removed; that depends on the system’s isolation design. For ordinary categories, the same separation prevents stale metadata from producing wrong answers. Record whether filtering happened before candidate generation, during traversal or scanning, or after retrieval, because those choices have different recall and cost behavior.