microsoft/semantic-kernel · error · ValueError

vectorEmbeddings cannot be null or empty in the vector_embed

Error message

vectorEmbeddings cannot be null or empty in the vector_embedding_policy.

What it means

Raised in AzureCosmosDBNoSQLMemoryStore.__init__ when vector_embedding_policy is None or its 'vectorEmbeddings' key maps to an empty list. The vectorEmbeddings array defines the embedding path, dimensions, and data type that Cosmos DB uses to index and search vectors. This check runs after the vectorIndexes check (line 48), so if both policies are invalid you see the vectorIndexes error first.

Source

Thrown at python/semantic_kernel/connectors/memory_stores/azure_cosmosdb_no_sql/azure_cosmosdb_no_sql_memory_store.py:51

    partition_key: str = None
    vector_embedding_policy: dict[str, Any] | None = None
    indexing_policy: dict[str, Any] | None = None
    cosmos_container_properties: dict[str, Any] | None = None

    def __init__(
        self,
        cosmos_client: CosmosClient,
        database_name: str,
        partition_key: str,
        vector_embedding_policy: dict[str, Any] | None = None,
        indexing_policy: dict[str, Any] | None = None,
        cosmos_container_properties: dict[str, Any] | None = None,
    ):
        """Initializes a new instance of the AzureCosmosDBNoSQLMemoryStore class."""
        if indexing_policy["vectorIndexes"] is None or len(indexing_policy["vectorIndexes"]) == 0:
            raise ValueError("vectorIndexes cannot be null or empty in the indexing_policy.")
        if vector_embedding_policy is None or len(vector_embedding_policy["vectorEmbeddings"]) == 0:
            raise ValueError("vectorEmbeddings cannot be null or empty in the vector_embedding_policy.")

        self.cosmos_client = cosmos_client
        self.database_name = database_name
        self.partition_key = partition_key
        self.vector_embedding_policy = vector_embedding_policy
        self.indexing_policy = indexing_policy
        self.cosmos_container_properties = cosmos_container_properties

    @override
    async def create_collection(self, collection_name: str) -> None:
        # Create the database if it already doesn't exist
        self.database = await self.cosmos_client.create_database_if_not_exists(id=self.database_name)

        # Create the collection if it already doesn't exist
        self.container = await self.database.create_container_if_not_exists(
            id=collection_name,
            partition_key=self.cosmos_container_properties["partition_key"],
            indexing_policy=self.indexing_policy,

View on GitHub (pinned to c028a0c7dc)

Solutions

Pass a vector_embedding_policy with a populated vectorEmbeddings list, e.g. {'vectorEmbeddings': [{'path': '/embedding', 'dataType': 'float32', 'dimensions': 1536, 'distanceFunction': 'cosine'}]}.
Make sure 'dimensions' matches the output dimension of your embedding model and 'path' matches the indexing_policy vectorIndexes path.
If building the policy from env/config, validate the dict shape before passing it to the constructor.
Migrate to the non-deprecated AzureCosmosDBNoSQLStore / Collection classes recommended by the @deprecated marker.

Example fix

// before
store = AzureCosmosDBNoSQLMemoryStore(
    cosmos_client=client,
    database_name='db',
    partition_key='/pk',
    vector_embedding_policy=None,  # triggers ValueError
    indexing_policy={'vectorIndexes': [{'path': '/embedding', 'type': 'diskANN'}]},
)
// after
vector_embedding_policy = {
    'vectorEmbeddings': [{
        'path': '/embedding',
        'dataType': 'float32',
        'dimensions': 1536,
        'distanceFunction': 'cosine',
    }]
}
store = AzureCosmosDBNoSQLMemoryStore(
    cosmos_client=client,
    database_name='db',
    partition_key='/pk',
    vector_embedding_policy=vector_embedding_policy,
    indexing_policy={'vectorIndexes': [{'path': '/embedding', 'type': 'diskANN'}]},
)

Defensive patterns

Strategy: validation

Validate before calling

def validate_vector_embedding_policy(policy: dict | None) -> None:
    if policy is None:
        raise ValueError('vector_embedding_policy must not be None')
    ve = policy.get('vectorEmbeddings')
    if ve is None or len(ve) == 0:
        raise ValueError('vector_embedding_policy["vectorEmbeddings"] must be non-empty')

validate_vector_embedding_policy(vector_embedding_policy)

Type guard

from typing import Any

def is_valid_vector_embedding_policy(policy: Any) -> bool:
    return (
        isinstance(policy, dict)
        and isinstance(policy.get('vectorEmbeddings'), list)
        and len(policy['vectorEmbeddings']) > 0
    )

Try / catch

try:
    store = AzureCosmosDBNoSQLMemoryStore(
        ..., vector_embedding_policy=vector_embedding_policy
    )
except ValueError as e:
    logging.error('Invalid vector embedding policy: %s', e)
    raise

Prevention

Never leave vector_embedding_policy as the default None.
Ensure dimensions match your embedding model output.
Validate the policy dict in config-loading code.
Align the embedding path with the indexing_policy vectorIndexes path.

When it happens

Trigger: Constructing AzureCosmosDBNoSQLMemoryStore with vector_embedding_policy=None (the default) or vector_embedding_policy={'vectorEmbeddings': []}. Because the None branch short-circuits with 'or' before subscripting, passing None does reach the ValueError rather than a TypeError here.

Common situations: Forgetting to pass vector_embedding_policy because it defaults to None. Providing a policy dict from a config file where the vectorEmbeddings array was stripped or never populated. Dimension/path mismatch between the embedding policy and the indexing policy.

Related errors

AI-assisted analysis of microsoft/semantic-kernel@c028a0c7dc (2026-08-13). Data as JSON: /api/errors/e574d8812aad014c. Report an issue: GitHub.