microsoft/semantic-kernel · error · VectorStoreModelValidationError

Field name '{IN_MEMORY_SCORE_KEY}' is reserved for internal

Error message

Field name '{IN_MEMORY_SCORE_KEY}' is reserved for internal use.

What it means

Thrown by _validate_data_model (in_memory.py:613-614) as VectorStoreModelValidationError when the data model defines a field named 'in_memory_search_score' (IN_MEMORY_SCORE_KEY). The collection writes vector search scores into each result record under that key, so a user field with the same name would collide and corrupt either the score or the data.

Source

Thrown at python/semantic_kernel/connectors/in_memory.py:614

        > - `max_filter_literal_collection_size=256`
        > - `max_filter_sequence_repeat_size=1024`
        > You can override these limits by passing them through `kwargs` or by setting them on the collection
        > instance after initialization.

        """
        super().__init__(
            record_type=record_type,
            definition=definition,
            collection_name=collection_name,
            embedding_generator=embedding_generator,
            **kwargs,
        )

    def _validate_data_model(self):
        """Check if the In Memory Score key is not used."""
        super()._validate_data_model()
        if IN_MEMORY_SCORE_KEY in self.definition.names:
            raise VectorStoreModelValidationError(f"Field name '{IN_MEMORY_SCORE_KEY}' is reserved for internal use.")

    @override
    async def _inner_delete(self, keys: Sequence[TKey], **kwargs: Any) -> None:
        for key in keys:
            self.inner_storage.pop(key, None)

    @override
    async def _inner_get(
        self, keys: Sequence[TKey] | None = None, options: GetFilteredRecordOptions | None = None, **kwargs: Any
    ) -> Any | OneOrMany[TModel] | None:
        if not keys:
            if options is not None:
                raise NotImplementedError("Get without keys is not yet implemented.")
            return None
        return [self.inner_storage[key] for key in keys if key in self.inner_storage]

    @override
    async def _inner_upsert(self, records: Sequence[Any], **kwargs: Any) -> Sequence[TKey]:

View on GitHub (pinned to c028a0c7dc)

Solutions

1. Rename the field to anything other than 'in_memory_search_score'.
2. If you need a score field, name it e.g. 'score', 'relevance', or 'my_score'.

Example fix

# before
class Rec(VectorStoreRecord):
    id: str
    in_memory_search_score: float   # reserved!
# after
class Rec(VectorStoreRecord):
    id: str
    relevance_score: float

RESERVED = 'in_memory_search_score'
def check_field_names(names: list[str]) -> None:
    if RESERVED in names:
        raise ValueError(f"field name '{RESERVED}' is reserved by InMemoryCollection")

from semantic_kernel.connectors.in_memory import IN_MEMORY_SCORE_KEY
def is_safe_field_name(name: str) -> bool:
    return name != IN_MEMORY_SCORE_KEY

try:
    collection = store.get_collection(record_type=Rec, definition=definition)
    await collection.create_collection()
except VectorStoreModelValidationError as e:
    # rename the colliding field and recreate the definition
    ...

When it happens

Trigger: Defining a VectorStoreRecordDefinition or pydantic model with a field/storage_name equal to 'in_memory_search_score' and creating/getting an InMemoryCollection for it; the validation runs during collection initialization.

Common situations: Choosing a generic score field name; copying a model schema from another store without checking reserved names; version upgrades that introduce the reserved key.

Related errors

• Vector field '{options.vector_property_name}' not found in t
• Chroma only supports one vector field, but {len(self.definit
• Attribute '{node.func.attr}' is not callable in filter expre
• Get without keys is not yet implemented.
• Filter string must be a lambda expression, e.g. 'lambda x: x

AI-assisted analysis of microsoft/semantic-kernel@c028a0c7dc (2026-08-13). Data as JSON: /api/errors/2fcab7d23742d875. Report an issue: GitHub.