lamindb 1.4.0__py3-none-any.whl → 1.5.1__py3-none-any.whl

lamindb/curators/core.py

@@ -0,0 +1,1534 @@
+"""Curator utilities.
+
+.. autosummary::
+   :toctree: .
+
+   Curator
+   SlotsCurator
+   CatVector
+   CatLookup
+   DataFrameCatManager
+
+"""
+
+from __future__ import annotations
+
+import copy
+import re
+from typing import TYPE_CHECKING, Any, Callable
+
+import lamindb_setup as ln_setup
+import numpy as np
+import pandas as pd
+import pandera
+from lamin_utils import colors, logger
+from lamindb_setup.core._docs import doc_args
+
+from lamindb.base.types import FieldAttr  # noqa
+from lamindb.models import (
+    Artifact,
+    Feature,
+    Record,
+    Run,
+    Schema,
+)
+from lamindb.models._from_values import _format_values
+from lamindb.models.artifact import (
+    data_is_anndata,
+    data_is_mudata,
+    data_is_spatialdata,
+)
+from lamindb.models.feature import parse_cat_dtype, parse_dtype
+
+from ..errors import InvalidArgument, ValidationError
+
+if TYPE_CHECKING:
+    from collections.abc import Iterable
+    from typing import Any
+
+    from anndata import AnnData
+    from mudata import MuData
+    from spatialdata import SpatialData
+
+    from lamindb.models.query_set import RecordList
+
+
+def strip_ansi_codes(text):
+    # This pattern matches ANSI escape sequences
+    ansi_pattern = re.compile(r"\x1b\[[0-9;]*m")
+    return ansi_pattern.sub("", text)
+
+
+class CatLookup:
+    """Lookup categories from the reference instance.
+
+    Args:
+        categoricals: A dictionary of categorical fields to lookup.
+        slots: A dictionary of slot fields to lookup.
+        public: Whether to lookup from the public instance. Defaults to False.
+
+    Example::
+
+        curator = ln.curators.DataFrameCurator(...)
+        curator.cat.lookup()["cell_type"].alveolar_type_1_fibroblast_cell
+
+    """
+
+    def __init__(
+        self,
+        categoricals: list[Feature] | dict[str, FieldAttr],
+        slots: dict[str, FieldAttr] = None,
+        public: bool = False,
+        sources: dict[str, Record] | None = None,
+    ) -> None:
+        slots = slots or {}
+        if isinstance(categoricals, list):
+            categoricals = {
+                feature.name: parse_dtype(feature.dtype)[0]["field"]
+                for feature in categoricals
+            }
+        self._categoricals = {**categoricals, **slots}
+        self._public = public
+        self._sources = sources
+
+    def __getattr__(self, name):
+        if name in self._categoricals:
+            registry = self._categoricals[name].field.model
+            if self._public and hasattr(registry, "public"):
+                return registry.public(source=self._sources.get(name)).lookup()
+            else:
+                return registry.lookup()
+        raise AttributeError(
+            f'"{self.__class__.__name__}" object has no attribute "{name}"'
+        )
+
+    def __getitem__(self, name):
+        if name in self._categoricals:
+            registry = self._categoricals[name].field.model
+            if self._public and hasattr(registry, "public"):
+                return registry.public(source=self._sources.get(name)).lookup()
+            else:
+                return registry.lookup()
+        raise AttributeError(
+            f'"{self.__class__.__name__}" object has no attribute "{name}"'
+        )
+
+    def __repr__(self) -> str:
+        if len(self._categoricals) > 0:
+            getattr_keys = "\n ".join(
+                [f".{key}" for key in self._categoricals if key.isidentifier()]
+            )
+            getitem_keys = "\n ".join(
+                [str([key]) for key in self._categoricals if not key.isidentifier()]
+            )
+            ref = "public" if self._public else "registries"
+            return (
+                f"Lookup objects from the {colors.italic(ref)}:\n "
+                f"{colors.green(getattr_keys)}\n "
+                f"{colors.green(getitem_keys)}\n"
+                'Example:\n    → categories = curator.lookup()["cell_type"]\n'
+                "    → categories.alveolar_type_1_fibroblast_cell\n\n"
+                "To look up public ontologies, use .lookup(public=True)"
+            )
+        else:  # pragma: no cover
+            return colors.warning("No fields are found!")
+
+
+CAT_MANAGER_DOCSTRING = """Manage categoricals by updating registries."""
+
+
+SLOTS_DOCSTRING = """Access sub curators by slot."""
+
+
+VALIDATE_DOCSTRING = """Validate dataset against Schema.
+
+Raises:
+    lamindb.errors.ValidationError: If validation fails.
+"""
+
+SAVE_ARTIFACT_DOCSTRING = """Save an annotated artifact.
+
+Args:
+    key: A path-like key to reference artifact in default storage, e.g., `"myfolder/myfile.fcs"`. Artifacts with the same key form a version family.
+    description: A description.
+    revises: Previous version of the artifact. Is an alternative way to passing `key` to trigger a new version.
+    run: The run that creates the artifact.
+
+Returns:
+    A saved artifact record.
+"""
+
+
+class Curator:
+    """Curator base class.
+
+    A `Curator` object makes it easy to validate, standardize & annotate datasets.
+
+    See:
+        - :class:`~lamindb.curators.DataFrameCurator`
+        - :class:`~lamindb.curators.AnnDataCurator`
+        - :class:`~lamindb.curators.MuDataCurator`
+        - :class:`~lamindb.curators.SpatialDataCurator`
+    """
+
+    def __init__(self, dataset: Any, schema: Schema | None = None):
+        self._artifact: Artifact = None  # pass the dataset as an artifact
+        self._dataset: Any = dataset  # pass the dataset as a UPathStr or data object
+        if isinstance(self._dataset, Artifact):
+            self._artifact = self._dataset
+            if self._artifact.otype in {
+                "DataFrame",
+                "AnnData",
+                "MuData",
+                "SpatialData",
+            }:
+                self._dataset = self._dataset.load(is_run_input=False)
+        self._schema: Schema | None = schema
+        self._is_validated: bool = False
+
+    @doc_args(VALIDATE_DOCSTRING)
+    def validate(self) -> bool | str:
+        """{}"""  # noqa: D415
+        pass  # pragma: no cover
+
+    @doc_args(SAVE_ARTIFACT_DOCSTRING)
+    def save_artifact(
+        self,
+        *,
+        key: str | None = None,
+        description: str | None = None,
+        revises: Artifact | None = None,
+        run: Run | None = None,
+    ) -> Artifact:
+        """{}"""  # noqa: D415
+        # Note that this docstring has to be consistent with the Artifact()
+        # constructor signature
+        pass  # pragma: no cover
+
+    def __repr__(self) -> str:
+        from lamin_utils import colors
+
+        if self._schema is not None:
+            # Schema might have different attributes
+            if hasattr(self._schema, "name") and self._schema.name:
+                schema_str = colors.italic(self._schema.name)
+            elif hasattr(self._schema, "uid"):
+                schema_str = colors.italic(f"uid={self._schema.uid}")
+            elif hasattr(self._schema, "id"):
+                schema_str = colors.italic(f"id={self._schema.id}")
+            else:
+                schema_str = colors.italic("unnamed")
+
+            # Add schema type info if available
+            if hasattr(self._schema, "otype") and self._schema.otype:
+                schema_str += f" ({self._schema.otype})"
+        else:
+            schema_str = colors.warning("None")
+
+        status_str = ""
+        if self._is_validated:
+            status_str = f", {colors.green('validated')}"
+        else:
+            status_str = f", {colors.yellow('unvalidated')}"
+
+        cls_name = colors.green(self.__class__.__name__)
+
+        # Get additional info based on curator type
+        extra_info = ""
+        if hasattr(self, "_slots") and self._slots:
+            # For SlotsCurator and its subclasses
+            slots_count = len(self._slots)
+            if slots_count > 0:
+                slot_names = list(self._slots.keys())
+                if len(slot_names) <= 3:
+                    extra_info = f", slots: {slot_names}"
+                else:
+                    extra_info = f", slots: [{', '.join(slot_names[:3])}... +{len(slot_names) - 3} more]"
+        elif (
+            cls_name == "DataFrameCurator"
+            and hasattr(self, "cat")
+            and hasattr(self.cat, "_categoricals")
+        ):
+            # For DataFrameCurator
+            cat_count = len(getattr(self.cat, "_categoricals", []))
+            if cat_count > 0:
+                extra_info = f", categorical_features={cat_count}"
+
+        artifact_info = ""
+        if self._artifact is not None:
+            artifact_uid = getattr(self._artifact, "uid", str(self._artifact))
+            short_uid = (
+                str(artifact_uid)[:8] + "..."
+                if len(str(artifact_uid)) > 8
+                else str(artifact_uid)
+            )
+            artifact_info = f", artifact: {colors.italic(short_uid)}"
+
+        return (
+            f"{cls_name}{artifact_info}(Schema: {schema_str}{extra_info}{status_str})"
+        )
+
+
+# default implementation for AnnDataCurator, MuDataCurator, and SpatialDataCurator
+class SlotsCurator(Curator):
+    """Curator for a dataset with slots.
+
+    Args:
+        dataset: The dataset to validate & annotate.
+        schema: A :class:`~lamindb.Schema` object that defines the validation constraints.
+
+    """
+
+    def __init__(
+        self,
+        dataset: Any,
+        schema: Schema,
+    ) -> None:
+        super().__init__(dataset=dataset, schema=schema)
+        self._slots: dict[str, DataFrameCurator] = {}
+
+        # used in MuDataCurator and SpatialDataCurator
+        # in form of {table/modality_key: var_field}
+        self._var_fields: dict[str, FieldAttr] = {}
+        # in form of {table/modality_key: categoricals}
+        self._cat_vectors: dict[str, dict[str, CatVector]] = {}
+
+    @property
+    @doc_args(SLOTS_DOCSTRING)
+    def slots(self) -> dict[str, DataFrameCurator]:
+        """{}"""  # noqa: D415
+        return self._slots
+
+    @doc_args(VALIDATE_DOCSTRING)
+    def validate(self) -> None:
+        """{}"""  # noqa: D415
+        for slot, curator in self._slots.items():
+            logger.info(f"validating slot {slot} ...")
+            curator.validate()
+
+    @doc_args(SAVE_ARTIFACT_DOCSTRING)
+    def save_artifact(
+        self,
+        *,
+        key: str | None = None,
+        description: str | None = None,
+        revises: Artifact | None = None,
+        run: Run | None = None,
+    ) -> Artifact:
+        """{}"""  # noqa: D415
+        if not self._is_validated:
+            self.validate()
+        if self._artifact is None:
+            if data_is_anndata(self._dataset):
+                self._artifact = Artifact.from_anndata(
+                    self._dataset,
+                    key=key,
+                    description=description,
+                    revises=revises,
+                    run=run,
+                )
+            if data_is_mudata(self._dataset):
+                self._artifact = Artifact.from_mudata(
+                    self._dataset,
+                    key=key,
+                    description=description,
+                    revises=revises,
+                    run=run,
+                )
+            elif data_is_spatialdata(self._dataset):
+                self._artifact = Artifact.from_spatialdata(
+                    self._dataset,
+                    key=key,
+                    description=description,
+                    revises=revises,
+                    run=run,
+                )
+            self._artifact.schema = self._schema
+            self._artifact.save()
+        cat_vectors = {}
+        for curator in self._slots.values():
+            for key, cat_vector in curator.cat._cat_vectors.items():
+                cat_vectors[key] = cat_vector
+        return annotate_artifact(  # type: ignore
+            self._artifact,
+            curator=self,
+            cat_vectors=cat_vectors,
+        )
+
+
+def check_dtype(expected_type) -> Callable:
+    """Creates a check function for Pandera that validates a column's dtype.
+
+    Args:
+        expected_type: String identifier for the expected type ('int', 'float', or 'num')
+
+    Returns:
+        A function that checks if a series has the expected dtype
+    """
+
+    def check_function(series):
+        if expected_type == "int":
+            is_valid = pd.api.types.is_integer_dtype(series.dtype)
+        elif expected_type == "float":
+            is_valid = pd.api.types.is_float_dtype(series.dtype)
+        elif expected_type == "num":
+            is_valid = pd.api.types.is_numeric_dtype(series.dtype)
+        return is_valid
+
+    return check_function
+
+
+# this is also currently used as DictCurator
+class DataFrameCurator(Curator):
+    # the example in the docstring is tested in test_curators_quickstart_example
+    """Curator for `DataFrame`.
+
+    Args:
+        dataset: The DataFrame-like object to validate & annotate.
+        schema: A :class:`~lamindb.Schema` object that defines the validation constraints.
+        slot: Indicate the slot in a composite curator for a composite data structure.
+
+    Example:
+
+        For simple example using a flexible schema, see :meth:`~lamindb.Artifact.from_df`.
+
+        Here is an example that enforces a minimal set of columns in the dataframe.
+
+        .. literalinclude:: scripts/curate_dataframe_minimal_errors.py
+            :language: python
+
+        Under-the-hood, this used the following schema.
+
+        .. literalinclude:: scripts/define_mini_immuno_schema_flexible.py
+            :language: python
+
+        Valid features & labels were defined as:
+
+        .. literalinclude:: scripts/define_mini_immuno_features_labels.py
+            :language: python
+    """
+
+    def __init__(
+        self,
+        dataset: pd.DataFrame | Artifact,
+        schema: Schema,
+        slot: str | None = None,
+    ) -> None:
+        super().__init__(dataset=dataset, schema=schema)
+        categoricals = []
+        features = []
+        feature_ids: set[int] = set()
+        if schema.flexible:
+            features += Feature.filter(name__in=self._dataset.keys()).list()
+            feature_ids = {feature.id for feature in features}
+        if schema.n > 0:
+            if schema._index_feature_uid is not None:
+                schema_features = [
+                    feature
+                    for feature in schema.members.list()
+                    if feature.uid != schema._index_feature_uid  # type: ignore
+                ]
+            else:
+                schema_features = schema.members.list()  # type: ignore
+            if feature_ids:
+                features.extend(
+                    feature
+                    for feature in schema_features
+                    if feature.id not in feature_ids  # type: ignore
+                )
+            else:
+                features.extend(schema_features)
+        else:
+            assert schema.itype is not None  # noqa: S101
+        pandera_columns = {}
+        if features or schema._index_feature_uid is not None:
+            # populate features
+            if schema.minimal_set:
+                optional_feature_uids = set(schema.optionals.get_uids())
+            for feature in features:
+                if schema.minimal_set:
+                    required = feature.uid not in optional_feature_uids
+                else:
+                    required = False
+                if feature.dtype in {"int", "float", "num"}:
+                    if isinstance(self._dataset, pd.DataFrame):
+                        dtype = (
+                            self._dataset[feature.name].dtype
+                            if feature.name in self._dataset.keys()
+                            else None
+                        )
+                    else:
+                        dtype = None
+                    pandera_columns[feature.name] = pandera.Column(
+                        dtype=None,
+                        checks=pandera.Check(
+                            check_dtype(feature.dtype),
+                            element_wise=False,
+                            error=f"Column '{feature.name}' failed dtype check for '{feature.dtype}': got {dtype}",
+                        ),
+                        nullable=feature.nullable,
+                        coerce=feature.coerce_dtype,
+                        required=required,
+                    )
+                else:
+                    pandera_dtype = (
+                        feature.dtype
+                        if not feature.dtype.startswith("cat")
+                        else "category"
+                    )
+                    pandera_columns[feature.name] = pandera.Column(
+                        pandera_dtype,
+                        nullable=feature.nullable,
+                        coerce=feature.coerce_dtype,
+                        required=required,
+                    )
+                if feature.dtype.startswith("cat"):
+                    # validate categoricals if the column is required or if the column is present
+                    if required or feature.name in self._dataset.keys():
+                        categoricals.append(feature)
+            if schema._index_feature_uid is not None:
+                # in almost no case, an index should have a pandas.CategoricalDtype in a DataFrame
+                # so, we're typing it as `str` here
+                index = pandera.Index(
+                    schema.index.dtype
+                    if not schema.index.dtype.startswith("cat")
+                    else str
+                )
+            else:
+                index = None
+            self._pandera_schema = pandera.DataFrameSchema(
+                pandera_columns,
+                coerce=schema.coerce_dtype,
+                strict=schema.maximal_set,
+                ordered=schema.ordered_set,
+                index=index,
+            )
+        self._cat_manager = DataFrameCatManager(
+            self._dataset,
+            columns_field=parse_cat_dtype(schema.itype, is_itype=True)["field"],
+            columns_names=pandera_columns.keys(),
+            categoricals=categoricals,
+            index=schema.index,
+            slot=slot,
+            maximal_set=schema.maximal_set,
+        )
+
+    @property
+    @doc_args(CAT_MANAGER_DOCSTRING)
+    def cat(self) -> DataFrameCatManager:
+        """{}"""  # noqa: D415
+        return self._cat_manager
+
+    def standardize(self) -> None:
+        """Standardize the dataset.
+
+        - Adds missing columns for features
+        - Fills missing values for features with default values
+        """
+        for feature in self._schema.members:
+            if feature.name not in self._dataset.columns:
+                if feature.default_value is not None or feature.nullable:
+                    fill_value = (
+                        feature.default_value
+                        if feature.default_value is not None
+                        else pd.NA
+                    )
+                    if feature.dtype.startswith("cat"):
+                        self._dataset[feature.name] = pd.Categorical(
+                            [fill_value] * len(self._dataset)
+                        )
+                    else:
+                        self._dataset[feature.name] = fill_value
+                    logger.important(
+                        f"added column {feature.name} with fill value {fill_value}"
+                    )
+                else:
+                    raise ValidationError(
+                        f"Missing column {feature.name} cannot be added because is not nullable and has no default value"
+                    )
+            else:
+                if feature.default_value is not None:
+                    if isinstance(
+                        self._dataset[feature.name].dtype, pd.CategoricalDtype
+                    ):
+                        if (
+                            feature.default_value
+                            not in self._dataset[feature.name].cat.categories
+                        ):
+                            self._dataset[feature.name] = self._dataset[
+                                feature.name
+                            ].cat.add_categories(feature.default_value)
+                    self._dataset[feature.name] = self._dataset[feature.name].fillna(
+                        feature.default_value
+                    )
+
+    def _cat_manager_validate(self) -> None:
+        self.cat.validate()
+
+        if self.cat._is_validated:
+            self._is_validated = True
+        else:
+            self._is_validated = False
+            raise ValidationError(self.cat._validate_category_error_messages)
+
+    @doc_args(VALIDATE_DOCSTRING)
+    def validate(self) -> None:
+        """{}"""  # noqa: D415
+        if self._schema.n > 0:
+            try:
+                # first validate through pandera
+                self._pandera_schema.validate(self._dataset)
+                # then validate lamindb categoricals
+                self._cat_manager_validate()
+            except pandera.errors.SchemaError as err:
+                self._is_validated = False
+                # .exconly() doesn't exist on SchemaError
+                raise ValidationError(str(err)) from err
+        else:
+            self._cat_manager_validate()
+
+    @doc_args(SAVE_ARTIFACT_DOCSTRING)
+    def save_artifact(
+        self,
+        *,
+        key: str | None = None,
+        description: str | None = None,
+        revises: Artifact | None = None,
+        run: Run | None = None,
+    ) -> Artifact:
+        """{}"""  # noqa: D415
+        if not self._is_validated:
+            self.validate()  # raises ValidationError if doesn't validate
+        if self._artifact is None:
+            self._artifact = Artifact.from_df(
+                self._dataset,
+                key=key,
+                description=description,
+                revises=revises,
+                run=run,
+                format=".csv" if key.endswith(".csv") else None,
+            )
+            self._artifact.schema = self._schema
+            self._artifact.save()
+        return annotate_artifact(  # type: ignore
+            self._artifact,
+            cat_vectors=self.cat._cat_vectors,
+        )
+
+
+class AnnDataCurator(SlotsCurator):
+    """Curator for `AnnData`.
+
+    Args:
+        dataset: The AnnData-like object to validate & annotate.
+        schema: A :class:`~lamindb.Schema` object that defines the validation constraints.
+
+    Example:
+
+        See :meth:`~lamindb.Artifact.from_anndata`.
+
+    """
+
+    def __init__(
+        self,
+        dataset: AnnData | Artifact,
+        schema: Schema,
+    ) -> None:
+        super().__init__(dataset=dataset, schema=schema)
+        if not data_is_anndata(self._dataset):
+            raise InvalidArgument("dataset must be AnnData-like.")
+        if schema.otype != "AnnData":
+            raise InvalidArgument("Schema otype must be 'AnnData'.")
+        self._slots = {
+            slot: DataFrameCurator(
+                (
+                    getattr(self._dataset, slot.strip(".T")).T
+                    if slot == "var.T"
+                    or (
+                        # backward compat
+                        slot == "var"
+                        and schema.slots["var"].itype not in {None, "Feature"}
+                    )
+                    else getattr(self._dataset, slot)
+                ),
+                slot_schema,
+                slot=slot,
+            )
+            for slot, slot_schema in schema.slots.items()
+            if slot in {"obs", "var", "var.T", "uns"}
+        }
+        if "var" in self._slots and schema.slots["var"].itype not in {None, "Feature"}:
+            logger.warning(
+                "auto-transposed `var` for backward compat, please indicate transposition in the schema definition by calling out `.T`: slots={'var.T': itype=bt.Gene.ensembl_gene_id}"
+            )
+            self._slots["var"].cat._cat_vectors["var_index"] = self._slots[
+                "var"
+            ].cat._cat_vectors.pop("columns")
+            self._slots["var"].cat._cat_vectors["var_index"]._key = "var_index"
+
+
+def _assign_var_fields_categoricals_multimodal(
+    modality: str | None,
+    slot_type: str,
+    slot: str,
+    slot_schema: Schema,
+    var_fields: dict[str, FieldAttr],
+    cat_vectors: dict[str, dict[str, CatVector]],
+    slots: dict[str, DataFrameCurator],
+) -> None:
+    """Assigns var_fields and categoricals for multimodal data curators."""
+    if modality is not None:
+        # Makes sure that all tables are present
+        var_fields[modality] = None
+        cat_vectors[modality] = {}
+
+    if slot_type == "var":
+        var_field = parse_cat_dtype(slot_schema.itype, is_itype=True)["field"]
+        if modality is None:
+            # This should rarely/never be used since tables should have different var fields
+            var_fields[slot] = var_field  # pragma: no cover
+        else:
+            # Note that this is NOT nested since the nested key is always "var"
+            var_fields[modality] = var_field
+    else:
+        obs_fields = slots[slot].cat._cat_vectors
+        if modality is None:
+            cat_vectors[slot] = obs_fields
+        else:
+            # Note that this is NOT nested since the nested key is always "obs"
+            cat_vectors[modality] = obs_fields
+
+
+class MuDataCurator(SlotsCurator):
+    """Curator for `MuData`.
+
+    Args:
+        dataset: The MuData-like object to validate & annotate.
+        schema: A :class:`~lamindb.Schema` object that defines the validation constraints.
+
+    Example:
+
+        .. literalinclude:: scripts/curate-mudata.py
+            :language: python
+            :caption: curate-mudata.py
+    """
+
+    def __init__(
+        self,
+        dataset: MuData | Artifact,
+        schema: Schema,
+    ) -> None:
+        super().__init__(dataset=dataset, schema=schema)
+        if not data_is_mudata(self._dataset):
+            raise InvalidArgument("dataset must be MuData-like.")
+        if schema.otype != "MuData":
+            raise InvalidArgument("Schema otype must be 'MuData'.")
+
+        for slot, slot_schema in schema.slots.items():
+            if ":" in slot:
+                modality, modality_slot = slot.split(":")
+                schema_dataset = self._dataset.__getitem__(modality)
+            else:
+                modality, modality_slot = None, slot
+                schema_dataset = self._dataset
+            if modality_slot == "var" and schema.slots[slot].itype not in {
+                None,
+                "Feature",
+            }:
+                logger.warning(
+                    "auto-transposed `var` for backward compat, please indicate transposition in the schema definition by calling out `.T`: slots={'var.T': itype=bt.Gene.ensembl_gene_id}"
+                )
+            self._slots[slot] = DataFrameCurator(
+                (
+                    getattr(schema_dataset, modality_slot.rstrip(".T")).T
+                    if modality_slot == "var.T"
+                    or (
+                        # backward compat
+                        modality_slot == "var"
+                        and schema.slots[slot].itype not in {None, "Feature"}
+                    )
+                    else getattr(schema_dataset, modality_slot)
+                ),
+                slot_schema,
+            )
+            _assign_var_fields_categoricals_multimodal(
+                modality=modality,
+                slot_type=modality_slot,
+                slot=slot,
+                slot_schema=slot_schema,
+                var_fields=self._var_fields,
+                cat_vectors=self._cat_vectors,
+                slots=self._slots,
+            )
+        self._columns_field = self._var_fields
+
+
+class SpatialDataCurator(SlotsCurator):
+    """Curator for `SpatialData`.
+
+    Args:
+        dataset: The SpatialData-like object to validate & annotate.
+        schema: A :class:`~lamindb.Schema` object that defines the validation constraints.
+
+    Example:
+
+        See :meth:`~lamindb.Artifact.from_spatialdata`.
+    """
+
+    def __init__(
+        self,
+        dataset: SpatialData | Artifact,
+        schema: Schema,
+        *,
+        sample_metadata_key: str | None = "sample",
+    ) -> None:
+        super().__init__(dataset=dataset, schema=schema)
+        if not data_is_spatialdata(self._dataset):
+            raise InvalidArgument("dataset must be SpatialData-like.")
+        if schema.otype != "SpatialData":
+            raise InvalidArgument("Schema otype must be 'SpatialData'.")
+
+        for slot, slot_schema in schema.slots.items():
+            split_result = slot.split(":")
+            if (len(split_result) == 2 and split_result[0] == "table") or (
+                len(split_result) == 3 and split_result[0] == "tables"
+            ):
+                if len(split_result) == 2:
+                    table_key, sub_slot = split_result
+                    logger.warning(
+                        f"please prefix slot {slot} with 'tables:' going forward"
+                    )
+                else:
+                    table_key, sub_slot = split_result[1], split_result[2]
+                slot_object = self._dataset.tables.__getitem__(table_key)
+                if sub_slot == "var" and schema.slots[slot].itype not in {
+                    None,
+                    "Feature",
+                }:
+                    logger.warning(
+                        "auto-transposed `var` for backward compat, please indicate transposition in the schema definition by calling out `.T`: slots={'var.T': itype=bt.Gene.ensembl_gene_id}"
+                    )
+                data_object = (
+                    getattr(slot_object, sub_slot.rstrip(".T")).T
+                    if sub_slot == "var.T"
+                    or (
+                        # backward compat
+                        sub_slot == "var"
+                        and schema.slots[slot].itype not in {None, "Feature"}
+                    )
+                    else getattr(slot_object, sub_slot)
+                )
+            elif len(split_result) == 1 or (
+                len(split_result) > 1 and split_result[0] == "attrs"
+            ):
+                table_key = None
+                if len(split_result) == 1:
+                    if split_result[0] != "attrs":
+                        logger.warning(
+                            f"please prefix slot {slot} with 'attrs:' going forward"
+                        )
+                        sub_slot = slot
+                        data_object = self._dataset.attrs[slot]
+                    else:
+                        sub_slot = "attrs"
+                        data_object = self._dataset.attrs
+                elif len(split_result) == 2:
+                    sub_slot = split_result[1]
+                    data_object = self._dataset.attrs[split_result[1]]
+                data_object = pd.DataFrame([data_object])
+            self._slots[slot] = DataFrameCurator(data_object, slot_schema, slot)
+            _assign_var_fields_categoricals_multimodal(
+                modality=table_key,
+                slot_type=sub_slot,
+                slot=slot,
+                slot_schema=slot_schema,
+                var_fields=self._var_fields,
+                cat_vectors=self._cat_vectors,
+                slots=self._slots,
+            )
+        self._columns_field = self._var_fields
+
+
+class CatVector:
+    """Vector with categorical values."""
+
+    def __init__(
+        self,
+        values_getter: Callable
+        | Iterable[str],  # A callable or iterable that returns the values to validate.
+        field: FieldAttr,  # The field to validate against.
+        key: str,  # The name of the vector to validate. Only used for logging.
+        values_setter: Callable | None = None,  # A callable that sets the values.
+        source: Record | None = None,  # The ontology source to validate against.
+        feature: Feature | None = None,
+        cat_manager: DataFrameCatManager | None = None,
+        subtype_str: str = "",
+        maximal_set: bool = True,  # whether unvalidated categoricals cause validation failure.
+    ) -> None:
+        self._values_getter = values_getter
+        self._values_setter = values_setter
+        self._field = field
+        self._key = key
+        self._source = source
+        self._organism = None
+        self._validated: None | list[str] = None
+        self._non_validated: None | list[str] = None
+        self._synonyms: None | dict[str, str] = None
+        self._subtype_str = subtype_str
+        self._subtype_query_set = None
+        self._cat_manager = cat_manager
+        self.feature = feature
+        self.records = None
+        self._maximal_set = maximal_set
+        if hasattr(field.field.model, "_name_field"):
+            label_ref_is_name = field.field.name == field.field.model._name_field
+        else:
+            label_ref_is_name = field.field.name == "name"
+        self.label_ref_is_name = label_ref_is_name
+
+    @property
+    def values(self):
+        """Get the current values using the getter function."""
+        if callable(self._values_getter):
+            return self._values_getter()
+        return self._values_getter
+
+    @values.setter
+    def values(self, new_values):
+        """Set new values using the setter function if available."""
+        if callable(self._values_setter):
+            self._values_setter(new_values)
+        else:
+            # If values_getter is not callable, it's a direct reference we can update
+            self._values_getter = new_values
+
+    @property
+    def is_validated(self) -> bool:
+        """Whether the vector is validated."""
+        # if nothing was validated, something likely is fundamentally wrong
+        # should probably add a setting `at_least_one_validated`
+        result = True
+        if len(self.values) > 0 and len(self.values) == len(self._non_validated):
+            result = False
+        # len(self._non_validated) != 0
+        #     if maximal_set is True, return False
+        #     if maximal_set is False, return True
+        # len(self._non_validated) == 0
+        #     return True
+        if len(self._non_validated) != 0:
+            if self._maximal_set:
+                result = False
+        return result
+
+    def _replace_synonyms(self) -> list[str]:
+        """Replace synonyms in the vector with standardized values."""
+        syn_mapper = self._synonyms
+        # replace the values in df
+        std_values = self.values.map(
+            lambda unstd_val: syn_mapper.get(unstd_val, unstd_val)
+        )
+        # remove the standardized values from self.non_validated
+        non_validated = [i for i in self._non_validated if i not in syn_mapper]
+        if len(non_validated) == 0:
+            self._non_validated = []
+        else:
+            self._non_validated = non_validated  # type: ignore
+        # logging
+        n = len(syn_mapper)
+        if n > 0:
+            syn_mapper_print = _format_values(
+                [f'"{k}" → "{v}"' for k, v in syn_mapper.items()], sep=""
+            )
+            s = "s" if n > 1 else ""
+            logger.success(
+                f'standardized {n} synonym{s} in "{self._key}": {colors.green(syn_mapper_print)}'
+            )
+        return std_values
+
+    def __repr__(self) -> str:
+        if self._non_validated is None:
+            status = "unvalidated"
+        else:
+            status = (
+                "validated"
+                if len(self._non_validated) == 0
+                else f"non-validated ({len(self._non_validated)})"
+            )
+
+        field_name = getattr(self._field, "name", str(self._field))
+        values_count = len(self.values) if hasattr(self.values, "__len__") else "?"
+        return f"CatVector(key='{self._key}', field='{field_name}', values={values_count}, {status})"
+
+    def _add_validated(self) -> tuple[list, list]:
+        """Save features or labels records in the default instance."""
+        from lamindb.models.save import save as ln_save
+
+        registry = self._field.field.model
+        field_name = self._field.field.name
+        model_field = registry.__get_name_with_module__()
+        filter_kwargs = get_current_filter_kwargs(
+            registry, {"organism": self._organism, "source": self._source}
+        )
+        values = [i for i in self.values if isinstance(i, str) and i]
+        if not values:
+            return [], []
+        # inspect the default instance and save validated records from public
+        if (
+            self._subtype_str != "" and "__" not in self._subtype_str
+        ):  # not for general filter expressions
+            self._subtype_query_set = registry.get(name=self._subtype_str).records.all()
+            values_array = np.array(values)
+            validated_mask = self._subtype_query_set.validate(  # type: ignore
+                values_array, field=self._field, **filter_kwargs, mute=True
+            )
+            validated_labels, non_validated_labels = (
+                values_array[validated_mask],
+                values_array[~validated_mask],
+            )
+            records = registry.from_values(
+                validated_labels, field=self._field, **filter_kwargs, mute=True
+            )
+        else:
+            existing_and_public_records = registry.from_values(
+                list(values), field=self._field, **filter_kwargs, mute=True
+            )
+            existing_and_public_labels = [
+                getattr(r, field_name) for r in existing_and_public_records
+            ]
+            # public records that are not already in the database
+            public_records = [r for r in existing_and_public_records if r._state.adding]
+            # here we check to only save the public records if they are from the specified source
+            # we check the uid because r.source and source can be from different instances
+            if self._source:
+                public_records = [
+                    r for r in public_records if r.source.uid == self._source.uid
+                ]
+            if len(public_records) > 0:
+                logger.info(f"saving validated records of '{self._key}'")
+                ln_save(public_records)
+                labels_saved_public = [getattr(r, field_name) for r in public_records]
+                # log the saved public labels
+                # the term "transferred" stresses that this is always in the context of transferring
+                # labels from a public ontology or a different instance to the present instance
+                if len(labels_saved_public) > 0:
+                    s = "s" if len(labels_saved_public) > 1 else ""
+                    logger.success(
+                        f'added {len(labels_saved_public)} record{s} {colors.green("from_public")} with {model_field} for "{self._key}": {_format_values(labels_saved_public)}'
+                    )
+                    # non-validated records from the default instance
+            non_validated_labels = [
+                i for i in values if i not in existing_and_public_labels
+            ]
+            validated_labels = existing_and_public_labels
+            records = existing_and_public_records
+
+        self.records = records
+        # validated, non-validated
+        return validated_labels, non_validated_labels
+
+    def _add_new(
+        self,
+        values: list[str],
+        df: pd.DataFrame | None = None,  # remove when all users use schema
+        dtype: str | None = None,
+        **create_kwargs,
+    ) -> None:
+        """Add new labels to the registry."""
+        from lamindb.models.save import save as ln_save
+
+        registry = self._field.field.model
+        field_name = self._field.field.name
+        non_validated_records: RecordList[Any] = []  # type: ignore
+        if df is not None and registry == Feature:
+            nonval_columns = Feature.inspect(df.columns, mute=True).non_validated
+            non_validated_records = Feature.from_df(df.loc[:, nonval_columns])
+        else:
+            if (
+                self._organism
+                and hasattr(registry, "organism")
+                and registry._meta.get_field("organism").is_relation
+            ):
+                # make sure organism record is saved to the current instance
+                create_kwargs["organism"] = _save_organism(name=self._organism)
+
+            for value in values:
+                init_kwargs = {field_name: value}
+                if registry == Feature:
+                    init_kwargs["dtype"] = "cat" if dtype is None else dtype
+                non_validated_records.append(registry(**init_kwargs, **create_kwargs))
+        if len(non_validated_records) > 0:
+            ln_save(non_validated_records)
+            model_field = colors.italic(registry.__get_name_with_module__())
+            s = "s" if len(values) > 1 else ""
+            logger.success(
+                f'added {len(values)} record{s} with {model_field} for "{self._key}": {_format_values(values)}'
+            )
+
+    def _validate(
+        self,
+        values: list[str],
+    ) -> tuple[list[str], dict]:
+        """Validate ontology terms using LaminDB registries."""
+        registry = self._field.field.model
+        field_name = self._field.field.name
+        model_field = f"{registry.__name__}.{field_name}"
+
+        kwargs_current = get_current_filter_kwargs(
+            registry, {"organism": self._organism, "source": self._source}
+        )
+
+        # inspect values from the default instance, excluding public
+        registry_or_queryset = registry
+        if self._subtype_query_set is not None:
+            registry_or_queryset = self._subtype_query_set
+        inspect_result = registry_or_queryset.inspect(
+            values, field=self._field, mute=True, from_source=False, **kwargs_current
+        )
+        non_validated = inspect_result.non_validated
+        syn_mapper = inspect_result.synonyms_mapper
+
+        # inspect the non-validated values from public (BioRecord only)
+        values_validated = []
+        if hasattr(registry, "public"):
+            public_records = registry.from_values(
+                non_validated,
+                field=self._field,
+                mute=True,
+                **kwargs_current,
+            )
+            values_validated += [getattr(r, field_name) for r in public_records]
+
+        # logging messages
+        if self._cat_manager is not None:
+            slot = self._cat_manager._slot
+        else:
+            slot = None
+        in_slot = f" in slot '{slot}'" if slot is not None else ""
+        slot_prefix = f".slots['{slot}']" if slot is not None else ""
+        non_validated_hint_print = (
+            f"curator{slot_prefix}.cat.add_new_from('{self._key}')"
+        )
+        non_validated = [i for i in non_validated if i not in values_validated]
+        n_non_validated = len(non_validated)
+        if n_non_validated == 0:
+            logger.success(
+                f'"{self._key}" is validated against {colors.italic(model_field)}'
+            )
+            return [], {}
+        else:
+            s = "" if n_non_validated == 1 else "s"
+            print_values = _format_values(non_validated)
+            warning_message = f"{colors.red(f'{n_non_validated} term{s}')} not validated in feature '{self._key}'{in_slot}: {colors.red(print_values)}\n"
+            if syn_mapper:
+                s = "" if len(syn_mapper) == 1 else "s"
+                syn_mapper_print = _format_values(
+                    [f'"{k}" → "{v}"' for k, v in syn_mapper.items()], sep=""
+                )
+                hint_msg = f'.standardize("{self._key}")'
+                warning_message += f"    {colors.yellow(f'{len(syn_mapper)} synonym{s}')} found: {colors.yellow(syn_mapper_print)}\n    → curate synonyms via: {colors.cyan(hint_msg)}"
+            if n_non_validated > len(syn_mapper):
+                if syn_mapper:
+                    warning_message += "\n    for remaining terms:\n"
+                warning_message += f"    → fix typos, remove non-existent values, or save terms via: {colors.cyan(non_validated_hint_print)}"
+                if self._subtype_query_set is not None:
+                    warning_message += f"\n    → a valid label for subtype '{self._subtype_str}' has to be one of {self._subtype_query_set.list('name')}"
+            logger.info(f'mapping "{self._key}" on {colors.italic(model_field)}')
+            logger.warning(warning_message)
+            if self._cat_manager is not None:
+                self._cat_manager._validate_category_error_messages = strip_ansi_codes(
+                    warning_message
+                )
+            return non_validated, syn_mapper
+
+    def validate(self) -> None:
+        """Validate the vector."""
+        # add source-validated values to the registry
+        self._validated, self._non_validated = self._add_validated()
+        self._non_validated, self._synonyms = self._validate(values=self._non_validated)
+
+        # always register new Features if they are columns
+        if self._key == "columns" and self._field == Feature.name:
+            self.add_new()
+
+    def standardize(self) -> None:
+        """Standardize the vector."""
+        registry = self._field.field.model
+        if not hasattr(registry, "standardize"):
+            return self.values
+        if self._synonyms is None:
+            self.validate()
+        # get standardized values
+        std_values = self._replace_synonyms()
+        # update non_validated values
+        self._non_validated = [
+            i for i in self._non_validated if i not in self._synonyms.keys()
+        ]
+        # remove synonyms since they are now standardized
+        self._synonyms = {}
+        # update the values with the standardized values
+        self.values = std_values
+
+    def add_new(self, **create_kwargs) -> None:
+        """Add new values to the registry."""
+        if self._non_validated is None:
+            self.validate()
+        if len(self._synonyms) > 0:
+            # raise error because .standardize modifies the input dataset
+            raise ValidationError(
+                "Please run `.standardize()` before adding new values."
+            )
+        self._add_new(
+            values=self._non_validated,
+            **create_kwargs,
+        )
+        # remove the non_validated values since they are now registered
+        self._non_validated = []
+
+
+class DataFrameCatManager:
+    """Manage categoricals by updating registries.
+
+    This class is accessible from within a `DataFrameCurator` via the `.cat` attribute.
+
+    If you find non-validated values, you have two options:
+
+    - new values found in the data can be registered via `DataFrameCurator.cat.add_new_from()` :meth:`~lamindb.curators.core.DataFrameCatManager.add_new_from`
+    - non-validated values can be accessed via `DataFrameCurator.cat.add_new_from()` :meth:`~lamindb.curators.core.DataFrameCatManager.non_validated` and addressed manually
+    """
+
+    def __init__(
+        self,
+        df: pd.DataFrame | Artifact,
+        columns_field: FieldAttr = Feature.name,
+        columns_names: Iterable[str] | None = None,
+        categoricals: list[Feature] | None = None,
+        sources: dict[str, Record] | None = None,
+        index: Feature | None = None,
+        slot: str | None = None,
+        maximal_set: bool = False,
+    ) -> None:
+        self._non_validated = None
+        self._index = index
+        self._artifact: Artifact = None  # pass the dataset as an artifact
+        self._dataset: Any = df  # pass the dataset as a UPathStr or data object
+        if isinstance(self._dataset, Artifact):
+            self._artifact = self._dataset
+            self._dataset = self._dataset.load(is_run_input=False)
+        self._is_validated: bool = False
+        self._categoricals = categoricals or []
+        self._non_validated = None
+        self._sources = sources or {}
+        self._columns_field = columns_field
+        self._validate_category_error_messages: str = ""
+        self._cat_vectors: dict[str, CatVector] = {}
+        self._slot = slot
+        self._maximal_set = maximal_set
+
+        if columns_names is None:
+            columns_names = []
+        if columns_field == Feature.name:
+            self._cat_vectors["columns"] = CatVector(
+                values_getter=columns_names,
+                field=columns_field,
+                key="columns" if isinstance(self._dataset, pd.DataFrame) else "keys",
+                source=self._sources.get("columns"),
+                cat_manager=self,
+                maximal_set=self._maximal_set,
+            )
+        else:
+            self._cat_vectors["columns"] = CatVector(
+                values_getter=lambda: self._dataset.columns,  # lambda ensures the inplace update
+                values_setter=lambda new_values: setattr(
+                    self._dataset, "columns", pd.Index(new_values)
+                ),
+                field=columns_field,
+                key="columns",
+                source=self._sources.get("columns"),
+                cat_manager=self,
+                maximal_set=self._maximal_set,
+            )
+        for feature in self._categoricals:
+            result = parse_dtype(feature.dtype)[
+                0
+            ]  # TODO: support composite dtypes for categoricals
+            key = feature.name
+            field = result["field"]
+            subtype_str = result["subtype_str"]
+            self._cat_vectors[key] = CatVector(
+                values_getter=lambda k=key: self._dataset[
+                    k
+                ],  # Capture key as default argument
+                values_setter=lambda new_values, k=key: self._dataset.__setitem__(
+                    k, new_values
+                ),
+                field=field,
+                key=key,
+                source=self._sources.get(key),
+                feature=feature,
+                cat_manager=self,
+                subtype_str=subtype_str,
+            )
+        if index is not None and index.dtype.startswith("cat"):
+            result = parse_dtype(index.dtype)[0]
+            field = result["field"]
+            key = "index"
+            self._cat_vectors[key] = CatVector(
+                values_getter=self._dataset.index,
+                field=field,
+                key=key,
+                feature=index,
+                cat_manager=self,
+            )
+
+    @property
+    def non_validated(self) -> dict[str, list[str]]:
+        """Return the non-validated features and labels."""
+        if self._non_validated is None:
+            raise ValidationError("Please run validate() first!")
+        return {
+            key: cat_vector._non_validated
+            for key, cat_vector in self._cat_vectors.items()
+            if cat_vector._non_validated and key != "columns"
+        }
+
+    @property
+    def categoricals(self) -> list[Feature]:
+        """The categorical features."""
+        return self._categoricals
+
+    def lookup(self, public: bool = False) -> CatLookup:
+        """Lookup categories.
+
+        Args:
+            public: If "public", the lookup is performed on the public reference.
+        """
+        return CatLookup(
+            categoricals=self._categoricals,
+            slots={"columns": self._columns_field},
+            public=public,
+            sources=self._sources,
+        )
+
+    def validate(self) -> bool:
+        """Validate variables and categorical observations."""
+        self._validate_category_error_messages = ""  # reset the error messages
+
+        validated = True
+        for key, cat_vector in self._cat_vectors.items():
+            logger.info(f"validating vector {key}")
+            cat_vector.validate()
+            validated &= cat_vector.is_validated
+        self._is_validated = validated
+        self._non_validated = {}  # type: ignore
+
+        if self._index is not None:
+            # cat_vector.validate() populates validated labels
+            # the index should become part of the feature set corresponding to the dataframe
+            if self._cat_vectors["columns"].records is not None:
+                self._cat_vectors["columns"].records.insert(0, self._index)  # type: ignore
+            else:
+                self._cat_vectors["columns"].records = [self._index]  # type: ignore
+
+        return self._is_validated
+
+    def standardize(self, key: str) -> None:
+        """Replace synonyms with standardized values.
+
+        Modifies the input dataset inplace.
+
+        Args:
+            key: The key referencing the column in the DataFrame to standardize.
+        """
+        if self._artifact is not None:
+            raise RuntimeError("can't mutate the dataset when an artifact is passed!")
+
+        if key == "all":
+            logger.warning(
+                "'all' is deprecated, please pass a single key from `.non_validated.keys()` instead!"
+            )
+            for k in self.non_validated.keys():
+                self._cat_vectors[k].standardize()
+        else:
+            self._cat_vectors[key].standardize()
+
+    def add_new_from(self, key: str, **kwargs):
+        """Add validated & new categories.
+
+        Args:
+            key: The key referencing the slot in the DataFrame from which to draw terms.
+            **kwargs: Additional keyword arguments to pass to create new records
+        """
+        if len(kwargs) > 0 and key == "all":
+            raise ValueError("Cannot pass additional arguments to 'all' key!")
+        if key == "all":
+            logger.warning(
+                "'all' is deprecated, please pass a single key from `.non_validated.keys()` instead!"
+            )
+            for k in self.non_validated.keys():
+                self._cat_vectors[k].add_new(**kwargs)
+        else:
+            self._cat_vectors[key].add_new(**kwargs)
+
+
+def get_current_filter_kwargs(registry: type[Record], kwargs: dict) -> dict:
+    """Make sure the source and organism are saved in the same database as the registry."""
+    db = registry.filter().db
+    source = kwargs.get("source")
+    organism = kwargs.get("organism")
+    filter_kwargs = kwargs.copy()
+
+    if isinstance(organism, Record) and organism._state.db != "default":
+        if db is None or db == "default":
+            organism_default = copy.copy(organism)
+            # save the organism record in the default database
+            organism_default.save()
+            filter_kwargs["organism"] = organism_default
+    if isinstance(source, Record) and source._state.db != "default":
+        if db is None or db == "default":
+            source_default = copy.copy(source)
+            # save the source record in the default database
+            source_default.save()
+            filter_kwargs["source"] = source_default
+
+    return filter_kwargs
+
+
+def get_organism_kwargs(
+    field: FieldAttr, organism: str | None = None, values: Any = None
+) -> dict[str, str]:
+    """Check if a registry needs an organism and return the organism name."""
+    registry = field.field.model
+    if registry.__base__.__name__ == "BioRecord":
+        import bionty as bt
+        from bionty._organism import is_organism_required
+
+        from ..models._from_values import get_organism_record_from_field
+
+        if is_organism_required(registry):
+            if organism is not None or bt.settings.organism is not None:
+                return {"organism": organism or bt.settings.organism.name}
+            else:
+                organism_record = get_organism_record_from_field(
+                    field, organism=organism, values=values
+                )
+                if organism_record is not None:
+                    return {"organism": organism_record.name}
+    return {}
+
+
+def annotate_artifact(
+    artifact: Artifact,
+    *,
+    curator: AnnDataCurator | SlotsCurator | None = None,
+    cat_vectors: dict[str, CatVector] | None = None,
+) -> Artifact:
+    from .. import settings
+    from ..models.artifact import add_labels
+
+    if cat_vectors is None:
+        cat_vectors = {}
+
+    # annotate with labels
+    for key, cat_vector in cat_vectors.items():
+        if (
+            cat_vector._field.field.model == Feature
+            or key == "columns"
+            or key == "var_index"
+        ):
+            continue
+        if len(cat_vector.records) > settings.annotation.n_max_records:
+            logger.important(
+                f"not annotating with {len(cat_vector.records)} labels for feature {key} as it exceeds {settings.annotation.n_max_records} (ln.settings.annotation.n_max_records)"
+            )
+            continue
+        add_labels(
+            artifact,
+            records=cat_vector.records,
+            feature=cat_vector.feature,
+            feature_ref_is_name=None,  # do not need anymore
+            label_ref_is_name=cat_vector.label_ref_is_name,
+            from_curator=True,
+        )
+
+    # annotate with inferred schemas aka feature sets
+    if artifact.otype == "DataFrame":
+        features = cat_vectors["columns"].records
+        if features is not None:
+            feature_set = Schema(
+                features=features, coerce_dtype=artifact.schema.coerce_dtype
+            )  # TODO: add more defaults from validating schema
+            if (
+                feature_set._state.adding
+                and len(features) > settings.annotation.n_max_records
+            ):
+                logger.important(
+                    f"not annotating with {len(features)} features as it exceeds {settings.annotation.n_max_records} (ln.settings.annotation.n_max_records)"
+                )
+                itype = parse_cat_dtype(artifact.schema.itype, is_itype=True)["field"]
+                feature_set = Schema(itype=itype, n=len(features))
+            artifact.feature_sets.add(
+                feature_set.save(), through_defaults={"slot": "columns"}
+            )
+    else:
+        for slot, slot_curator in curator._slots.items():
+            # var_index is backward compat (2025-05-01)
+            name = (
+                "var_index"
+                if (slot == "var" and "var_index" in slot_curator.cat._cat_vectors)
+                else "columns"
+            )
+            features = slot_curator.cat._cat_vectors[name].records
+            if features is None:
+                logger.warning(f"no features found for slot {slot}")
+                continue
+            itype = parse_cat_dtype(artifact.schema.slots[slot].itype, is_itype=True)[
+                "field"
+            ]
+            feature_set = Schema(features=features, itype=itype)
+            if (
+                feature_set._state.adding
+                and len(features) > settings.annotation.n_max_records
+            ):
+                logger.important(
+                    f"not annotating with {len(features)} features for slot {slot} as it exceeds {settings.annotation.n_max_records} (ln.settings.annotation.n_max_records)"
+                )
+                feature_set = Schema(itype=itype, n=len(features))
+            artifact.feature_sets.add(
+                feature_set.save(), through_defaults={"slot": slot}
+            )
+
+    slug = ln_setup.settings.instance.slug
+    if ln_setup.settings.instance.is_remote:  # pdagma: no cover
+        logger.important(f"go to https://lamin.ai/{slug}/artifact/{artifact.uid}")
+    return artifact
+
+
+# TODO: need this function to support mutli-value columns
+def _flatten_unique(series: pd.Series[list[Any] | Any]) -> list[Any]:
+    """Flatten a Pandas series containing lists or single items into a unique list of elements."""
+    result = set()
+
+    for item in series:
+        if isinstance(item, list):
+            result.update(item)
+        else:
+            result.add(item)
+
+    return list(result)
+
+
+def _save_organism(name: str):
+    """Save an organism record."""
+    import bionty as bt
+
+    organism = bt.Organism.filter(name=name).one_or_none()
+    if organism is None:
+        organism = bt.Organism.from_source(name=name)
+        if organism is None:
+            raise ValidationError(
+                f'Organism "{name}" not found from public reference\n'
+                f'      → please save it from a different source: bt.Organism.from_source(name="{name}", source).save()'
+                f'      → or manually save it without source: bt.Organism(name="{name}").save()'
+            )
+        organism.save()
+    return organism