Skip to main content
Builds the index using the specified training configuration. Required before efficient querying. Prior to calling this, all queries will be conducted using encrypted exhaustive search. After, they will be conducted using encrypted ANN search. 
def train(self,
          n_lists: int = None,
          batch_size: int = None,
          max_iters: int = None,
          tolerance: float = None,
          max_memory: int = None,
          *,
          index_key: bytes = None,
          user_id: bytes = None)

​
Parameters

ParameterTypeDefaultDescription
n_listsintNone(Optional) Number of inverted index lists to create in the index. When None, auto-determines based on the number of vectors in the index (equivalent to 0).
batch_sizeintNone(Optional) Size of each batch for training. When None, defaults to 0 (auto — the optimal batch size is selected automatically).
max_itersintNone(Optional) Maximum number of iterations for training. When None, defaults to 100.
tolerancefloatNone(Optional) Convergence tolerance for training. When None, defaults to 1e-6.
max_memoryintNone(Optional) Maximum memory to use for training. When None, defaults to 0 (no limit).
index_keybytesNone(Optional, keyword-only) Override the per-operation index key. See Per-operation key override.
user_idbytesNone(Optional, keyword-only) 16-byte RBAC user identifier. See Per-operation key override.
There must be at least 2 * n_lists or 10,000 (whichever is greater) vector embeddings in the index prior to calling this function.

​
Exceptions

  • Raises an exception if the index was not created or loaded yet.
  • Raises an exception if there are not enough vector embeddings in the index for training (must be at least 2 * n_lists).
  • Raises an exception if the index could not be trained.

​
Example Usage

# Load index
index = client.load_index(
    index_name=index_name, 
    index_key=index_key
)

# Train the index with custom settings
index.train(
    batch_size=128, 
    max_iters=10, 
    tolerance=1e-4
)

# Train with default settings (auto-selected configuration)
index.train()

​
Per-operation key override

The calls above reuse the key supplied at create_index() / load_index(). You may instead pass index_key= (and user_id= for an RBAC user) to override the per-operation key. This is required in stateless/service deployments that reload the index per request: 
index.train(index_key=index_key)
Training requires write permission. An RBAC user without a write wrap cannot train the index.