Repository navigation
fix: find Portuguese houses under the national register's street names - #224
Conversation
… street names Portugal's houses come from the CTT register, which writes "R FARIAS" for the map's "Rua dos Farias" (types and titles abbreviated, the particle after the type dropped). Search now tries that spelling when the stated road has no such house, SuggestAddress does too for Portugal, and a number libpostal leaves before a Portuguese "Avenida" becomes the house number. global_osm_pt: SearchAddress 22 -> 46/100, SuggestAddress 22 -> 46/100. Refs #222.
Speed, full index: no slowdown over 10%DetailsSearchAddressTime in ms over 1995 cases, each the fastest of 3 requests.
SuggestAddressTime in ms over 1996 checks, each the fastest of 3 requests.
ReverseGeocodeTime in ms over 1994 cases, each the fastest of 3 requests.
Fails when the total, p50, p90 or p99 of any table is more than 10% slower. Runs of one binary differ by up to 5%; the max is one request and only reported. Base |
Gold check, full index: no regressions
Pass rate by countryPassing checks out of those checked, baseline → candidate, for the 1 expected country where any RPC changed. A case that expects several countries has its own row. Counts are checks, not independent locations.
Runs
SearchAddress: changes by cohort
Only cohorts with a change and kinds that occur are shown. gained/lost: a result appeared/disappeared; finer/coarser: resolution level; moved: >1 km; nudged: <1 km but visible at 3 decimals. SuggestAddress
Newly passing (verify before relabelling) (48)
SearchAddress: changed cases (27)finer (27)
Landmark checks (head)
|
Merging this PR will not alter performance
Comparing Footnotes
|
Summary
Portuguese address search and autocomplete now return the house instead of stopping at the street. On the
global_osm_ptpanel, SearchAddress and SuggestAddress both go from 22/100 to 46/100 on a local full index. Refs #222 (Portugal rows only; the issue tracks many markets).Every Portuguese house in the index comes from Overture's
pt/countrywideaddresses, which use the national register (CTT) spelling:R FARIASfor the map'sRua dos Farias, with the type and titles abbreviated and the particle after the type dropped. Queries matched the map street, which has no houses, so they answered at street level. A second defect: libpostal leaves a number beforeAvenidainside the road.Changes
Santa/SantoasS, particle after the type dropped).straat/streetswap.avenidain the road becomes the house number. Not applied elsewhere, since6 Avenidais a street name in Guatemala.docs/evaluation/triage/portugal.md.The builder and index are unchanged, so the builder hash is unchanged.
Review focus
Check
ctt_roadinsrc/matching.rsfor abbreviations that would over-match, and the Portugal-only gate insrc/index/suggest.rs.Results
Local full index
74a1eec034f5(builder9d4006a4eec15abb), base7a02601, same evaluator, models and reference data:global_osm_ptSearchAddress 22 to 46/100 (tuning groups 9 to 19/39, held-out 2 to 4/11; held-out groups were scored, not inspected).4 Rua da Imprensa à Estrelanow resolves to house 4 on the same street, about 100 m from the old street point.Checks run:
cargo fmt --check,clippy -D warningson all targets,cargo test, and the synthetic tests including--ignored. The new Portugal search and suggest cases fail on base and pass on head.tests/rpc.rsrpc_contractsfails on both base and head at the same CDG landmark check, because the local index has no answer for "CDG". It is unrelated to this change.Remaining Portuguese tuning failures (20 groups): 18 numbers absent from the index (source coverage), and 2 name mismatches not bridged (
Virgílio/Vergíliospelling, theGNRacronym).