Search space and time.

One index, four modalities

A query like “every structure within 2 km of this polygon that changed between March and July, ranked by similarity to a reference image” touches geometry, time, imagery, and semantics at once. On a conventional stack that is four systems and a join written in application code. In Lucenia it is one query, and relevance is computed where the index statistics live.

CONVENTIONAL                                LUCENIA

Elasticsearch   full-text  ─┐               ┌─────────────────────────┐
Qdrant          vectors    ─┤ application   │ single Lucene index     │
PostGIS         geometry   ─┤ join layer    │ text · vector · geo · xy│
GDAL + glue     rasters    ─┘               └─────────────────────────┘
                                                        │
4 systems, 4 copies of the data,            one query, one result set
4 accreditation packages                    one accreditation package

Ranking is fused in the query planner, not in a re-ranking pass over candidates fetched from three other systems.

// one request — text relevance, vector similarity, geometry, and time
POST /scenes/_search
{
  "query": {
    "bool": {
      "must": [
        { "knn":   { "embedding": { "vector": [0.12, …], "k": 50 } } },
        { "match": { "annotation": "construction crane" } }
      ],
      "filter": [
        { "geo_shape": { "footprint": { "shape": { "type": "polygon", … },
                                        "relation": "intersects" } } },
        { "range": { "acquired": { "gte": "2026-03-01", "lte": "2026-07-31" } } }
      ]
    }
  }
}

Spatial is not a plugin

Real geometry — points, lines, circles, polygons — with the full predicate set (INTERSECTS DISJOINT WITHIN CONTAINS) in both geographic and Cartesian coordinates. The Cartesian side makes imagery addressable: detections on a raster are query targets exactly like map geometry, and geometry participates in ranking directly alongside BM25 and k-NN scores.

Don’t take the diagram’s word for it

imagecat is a live image catalog running on Lucenia: 144 Sentinel-2 scenes, searchable in natural language — “show me orchards in the Central Valley” — with multimodal semantic retrieval, the reasoning behind each match exposed, and governance filtering applied at query time.

Query the archive without ingesting it

Lucenia reads Parquet through a custom Lucene codec: a row becomes a document, a column becomes a field, a file becomes a segment. Mount files where they already sit in object storage and query them in place — then promote a slice to a full index when it earns full-text and vector search. Cold data costs storage, not ingest.

Your data never leaves your trust boundary

Four deployment models, one platform, one API. The difference is who runs the infrastructure — never where your data lives.

Where one engine replaces four

Geospatial intelligence, earth observation, catastrophe modeling, public safety, critical infrastructure — workloads where the data is spatial, multimodal, and cannot leave the building.

Bring us a query your stack can’t answer

The difference is easiest to see on a query that touches geometry, imagery, and text at once.