RTREE index

RTREE is a spatial index designed for GEOMETRY fields. If your collection stores geographic or geometric shapes (points, lines, polygons) in WKT (Well‑Known Text) format and you frequently filter by where a shape is (inside, intersects, etc.), RTREE can make those queries much faster.

Typical use cases:

• Geo apps: “Find stores inside this polygon” (delivery zones, service areas)
• Routing / mapping: “Return assets intersecting this road segment”
• Operational analytics: “Show devices within this bounding box”

Note: RTREE requires backend support for GEOMETRY + RTREE (Milvus 2.6.4+).

How RTREE works

RTREE accelerates spatial filtering using a two‑phase approach:

1. Index build (organize shapes)

   • For every geometry object, the system computes a Minimum Bounding Rectangle (MBR) — the smallest rectangle that fully contains the object.
   • Nearby MBRs are grouped into larger rectangles, forming a tree (leaf nodes → internal nodes → root).

2. Query time (prune + verify)

   • The query geometry is also converted into an MBR.
   • The tree is traversed from the root and branches whose rectangles do not intersect the query MBR are skipped.
   • Only the remaining candidates are checked with the exact spatial predicate (contains / intersects / within, etc.).

This is why RTREE is fast: it avoids scanning every row by quickly discarding geometry that cannot match.

Create an RTREE index

You create RTREE on a GEOMETRY field in your collection schema (for example, a field named geo).

from dodil import Client
from dodil.vbase import VBaseConfig
 
c = Client(
    service_account_id="...",
    service_account_secret="...",
)
 
vbase = c.vbase.connect(
    VBaseConfig(
        host="vbase-db-<id>.infra.dodil.cloud",
        port=443,
        scheme="https",
        db_name="db_<id>",
    )
)
 
# Create an RTREE index on a GEOMETRY field.
# (The params dict is empty because RTREE does not require extra build parameters.)
vbase.create_index(
    collection_name="geo_demo",
    field_name="geo",
    index_type="RTREE",
    index_name="rtree_geo",
    params={},
)

After creating the index, load the collection (if your workflow requires explicit loading):

vbase.load_collection("geo_demo")

Query with spatial filters

RTREE is used automatically when your filter targets the indexed GEOMETRY field.

Example 1: filter-only query

Find shapes fully contained inside a polygon:

filter_expr = "ST_CONTAINS(geo, 'POLYGON ((0 0, 10 0, 10 10, 0 10, 0 0))')"
 
rows = vbase.query(
    collection_name="geo_demo",
    filter=filter_expr,
    output_fields=["id", "geo"],
    limit=10,
)
 
print(rows)

Example 2: vector search + spatial filter

If your collection also has a vector field (for example vec), you can combine semantic similarity with spatial constraints.

query_vec = [[0.1, 0.2, 0.3, 0.4, 0.5]]
filter_expr = "ST_INTERSECTS(geo, 'LINESTRING (1 1, 2 2)')"
 
hits = vbase.search(
    collection_name="geo_demo",
    data=query_vec,
    limit=5,
    filter=filter_expr,
    output_fields=["id", "geo"],
)
 
print(hits)

If an RTREE index exists, VBase can prune most candidates before applying the exact spatial predicate.

Drop the index

When you no longer need an index, drop it by name:

vbase.drop_index(
    collection_name="geo_demo",
    index_name="rtree_geo",
)

Practical tips

• WKT input matters: store geometry consistently (same coordinate order and units) so filters behave as expected.
• RTREE is for filters: it speeds up spatial predicates (contains/intersects/within), not vector similarity itself.
• Expect best gains when filters are selective: the more a spatial filter narrows the search area, the bigger the speedup.