geosearch-index

Prebuilt serving artifacts for a multilingual toponym search engine: enough to answer queries with no database of any kind.

file	what it is
bm25_index.npz	character-n-gram BM25 index over 1,190,895 name strings, plus the geonameid groups each name resolves to
places.parquet	1,304,659 places — the fields a result needs (name, country, population, feature code, coordinates)
descriptions.parquet	1,304,659 candidate documents for the reranker
manifest.json	what this build is, and what it was built from

No pickle. The index is plain numpy arrays loaded with allow_pickle=False. An earlier format was a pickled object graph, which is arbitrary code execution on load — acceptable for a file you just wrote yourself, not for one fetched from a public repo.

These four files are one set

manifest.json carries a build_id, and the two Parquet files are stamped with it. That is checked at load, because the failure it prevents is silent: a places table missing a geonameid the index can retrieve does not raise, it just drops that place from every result it should have won.

build_id  d2661ea7-2d41-4ec2-a879-d400579ac766
built_at  2026-08-07T08:39:20+00:00

Scope

countries	RU, US, TR, CN
languages	ru, en, tr, zh
excluded feature codes	PPLH, PPLQ, PPLW, PPLX

Scope is recorded because it decides what is in the corpus, not merely how it is stored — an index built for two countries and served under a config naming four cannot return the missing two, and nothing downstream would notice.

Retrieval is name-level, not place-level

One BM25 document per unique name string, backed by every geonameid carrying it. So "Moscow" is a single document covering the capital and every hamlet sharing the spelling, and they all inherit the same retriever score — homonyms tie at retrieval time by construction. Disambiguation is deferred to the population tiebreak and the reranker, not solved by the retriever.

IDF is pinned to 1, making the score pure term-frequency saturation over shared character n-grams: deliberately tolerant of transliteration and fuzzy spelling across ru, en, tr, zh, at the cost of rare-term weighting.

Checksums

bm25_index.npz  8e3e27f14c9487dd3b21c247ddc0ecebf47ba8722b3c2e598366e59a35e28dd3
places.parquet  e80fda97872e55a7e9ab33e990b0ec23f0dddd21901dfec5704c38b75e07b999
descriptions.parquet  6866300937684468e5d1b8d56009c0c358c2e2eed104cfc1363c358fc1694e96

Data licence and attribution

Derived from GeoNames, licensed CC BY 4.0.

Modifications made to the source data:

• filtered to populated places only (feature_class = 'P')
• restricted to RU, US, TR, CN
• restricted to name variants in ru, en, tr, zh
• dropped feature codes PPLH, PPLQ, PPLW, PPLX
• name variants grouped per place and joined into single document strings