Skip to content

Source: milvus

hms-import milvus --uri URI_OR_FILE --collection NAME [--vector-field F] [--text-field F] [--namespace-field F] [--partition P ...] [--created-at-field F] [--updated-at-field F] [--page-size N]

Reads a collection through pymilvus.MilvusClient: a Milvus Lite database file, a Milvus server (http://host:19530) or Zilliz Cloud. Verified against real Milvus Lite (pymilvus 3.0, milvus-lite) including partitions, a VARCHAR primary key and dynamic fields. A Milvus server and Zilliz Cloud use the same client calls but were not run here (no server available); treat them as unverified.

Mapping

  • Id: the primary key, as a string.
  • Vector: the single FLOAT_VECTOR field (or --vector-field). Binary, fp16, bf16 and sparse vectors are not supported; the run fails with the field list.
  • Text: --text-field, else the first field named text, content, document, chunk, body or page_content.
  • Metadata: every other field, including dynamic fields (returned flat by output_fields=["*"]). Nested JSON is preserved.
  • Namespace: the partition name (_default becomes the empty namespace), or the value of --namespace-field when given. --partition restricts which partitions are read.
  • Timestamps: --created-at-field / --updated-at-field (copied; the field also stays in metadata).

Reading and resume

query_iterator, one partition at a time in primary-key order, --page-size rows per round trip (default 1000, max 16384). Seekable: the cursor is [partition index, last primary key] and resume uses a pk > last filter inside the iterator. Rows written during the import may or may not be seen, as with any live collection.

Credentials

$MILVUS_TOKEN (or --token), $MILVUS_USER / $MILVUS_PASSWORD. A password inside the URI is also masked. None appear in reports, checkpoints or errors.