UpdateIndex

Update an existing vector index on a descriptor class. Requires indexes.create permission. Ordinary scalar/text/array index kinds do not support this command.

An exact node target is its type and class, for example {"target":"image","class":"Photos"}.

For a connection, class names the source class, connection class, and destination class. target supplies the corresponding endpoint types.

{
  "class": {"src": "People", "connection": "AppearsIn", "dst": "Photos"},
  "target": {"src": "entity", "dst": "image"}
}

Metadata mutations require all five values for a connection. Class names are nonempty, cannot begin with _, and are matched without case sensitivity. Index selectors also include property and kind. With GetIndexes and GetAggregates you can supply only the fields you want to match, such as {"class":{"connection":"AppearsIn"},"target":{"dst":"image"}}.

Supply property: "_blob" and the index kind (flat, ivf, hnsw, or acorn) with at least one action.

  • params completely replaces the vector configuration. Existing nonmissing vectors must satisfy the replacement dimensions. An invalid replacement rolls back. See Descriptor classes and vector indexes for the parameters of each kind.

  • metadata replaces customer metadata. An empty object clears it. Metadata changes do not rebuild vectors.

  • build_index: true runs the existing build/training barrier while retaining the selected index’s lazy mode.

The durable index identity is preserved. Construction changes use a fresh generation; query tuning preserves compatible generations. A successful response contains status: 0. Removing and recreating an index instead allocates a new identity and cannot publish work prepared for the old index.

[{"UpdateIndex":{"target":"descriptor","class":"Faces","property":"_blob",
                 "kind":"hnsw","metadata":{"model":"face-v2"},
                 "build_index":true}}]