Tokyo points-of-interest with map-driven bounding-box search combined with Japanese full-text and embedding queries.
On first visit the demo fetches the UniDic dictionary (~52 MB) and the multilingual MiniLM embedding model (~120 MB). Both are cached for offline reuse and shared with the basic sample.
Text input is wrapped in the unified DSL. Examples:
東京 title:浅草寺
embedding:"夜景がきれいなスポット".
With Filter by current map view on, the demo emits
+(query) +location:geo_bbox(...) so a document
must match both the text and the current map viewport.
Adds location:geo_bbox(min_lat, min_lon, max_lat, max_lon)
(derived from the current Leaflet bounds) to your query.
Live view of the map state and the DSL string the demo
sends to index.search(). Updates while you
pan or zoom.
————Delete persisted state if you have stale data from a previous demo version. Reloads the page after clearing.