The four queries that did not yield a top-1 match. Each is informative.
| Query | First-relevant rank | Top-1 returned | Note |
|---|---|---|---|
SwiftUI View | not in top 10 | apple-docs://clockkit/swiftui-templates | Two-word query. View token bound to swiftui-templates URI dominates over the canonical apple-docs://swiftui/view. Probable cause: the query text SwiftUI View ranks pages with both tokens highly, and clockkit/swiftui-templates contains both, while swiftui/view contains them less prominently in the indexed body. Realistic regression class for users typing two-word lookups for framework-rooted concepts. |
Combine Publisher | 2 | apple-docs://combine | Top-1 is the framework root; the canonical combine/publisher page is at rank 2. Defensible result for a two-word query naming both the framework and the protocol; an LLM consumer reading top-2 sees what it needs. |
CoreData | 3 | apple-docs://foundation/cocoaerror/code/coredata | Top-1 is the CocoaError.Code.coreData enum case, not the Core Data framework root. Probable cause: framework_aliases table maps coredata → CoreData but the framework root isn't winning the BM25F tie. The two preceding hits are deep API pages that contain the literal token coredata. |
MapKit | 2 | apple-docs://mapkitjs/mapkit | Top-1 is the JavaScript variant (mapkitjs) instead of the native mapkit framework. Probable cause: mapkitjs/mapkit contains the token in both framework and path positions; native mapkit only in the framework position. Source-authority weights should bias toward native; they don't here. |
None of these is a regression vs the brew baseline (the paired pilot run on 2026-05-20 had the same outcomes); they are pre-existing ranking artefacts of the current BM25F + RRF configuration. Each is a candidate for a per-class follow-up evaluation per docs/design/search-quality-eval.md §14.2 (specifically class B framework-root for SwiftUI View, Combine Publisher, MapKit; class C acronym/synonym for CoreData).