linkml-store 0.1.13__py3-none-any.whl → 0.1.14__py3-none-any.whl

linkml_store/inference/implementations/rag_inference_engine.py

@@ -0,0 +1,145 @@
+import logging
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import yaml
+from llm import get_key
+
+from linkml_store.api.collection import OBJECT, Collection
+from linkml_store.inference.inference_config import Inference, InferenceConfig, LLMConfig
+from linkml_store.inference.inference_engine import InferenceEngine
+
+logger = logging.getLogger(__name__)
+
+SYSTEM_PROMPT = """
+You are a {llm_config.role}, your task is to inference the YAML
+object output given the YAML object input. I will provide you
+with a collection of examples that will provide guidance both
+on the desired structure of the response, as well as the kind
+of content.
+
+You should return ONLY valid YAML in your response.
+"""
+
+
+@dataclass
+class RAGInferenceEngine(InferenceEngine):
+    """
+    AI Retrieval Augmented Generation (RAG) based predictor.
+
+
+    >>> from linkml_store.api.client import Client
+    >>> from linkml_store.utils.format_utils import Format
+    >>> from linkml_store.inference.inference_config import LLMConfig
+    >>> client = Client()
+    >>> db = client.attach_database("duckdb", alias="test")
+    >>> db.import_database("tests/input/countries/countries.jsonl", Format.JSONL, collection_name="countries")
+    >>> db.list_collection_names()
+    ['countries']
+    >>> collection = db.get_collection("countries")
+    >>> features = ["name"]
+    >>> targets = ["code", "capital", "continent", "languages"]
+    >>> llm_config = LLMConfig(model_name="gpt-4o-mini",)
+    >>> config = InferenceConfig(target_attributes=targets, feature_attributes=features, llm_config=llm_config)
+    >>> ie = RAGInferenceEngine(config=config)
+    >>> ie.load_and_split_data(collection)
+    >>> ie.initialize_model()
+    >>> prediction = ie.derive({"name": "Uruguay"})
+    >>> prediction.predicted_object
+    {'capital': 'Montevideo', 'code': 'UY', 'continent': 'South America', 'languages': ['Spanish']}
+
+    """
+
+    classifier: Any = None
+    encoders: dict = None
+    _model: "llm.Model" = None  # noqa: F821
+
+    rag_collection: Collection = None
+
+    def __post_init__(self):
+        if not self.config:
+            self.config = InferenceConfig()
+        if not self.config.llm_config:
+            self.config.llm_config = LLMConfig()
+
+    @property
+    def model(self) -> "llm.Model":  # noqa: F821
+        import llm
+
+        if self._model is None:
+            self._model = llm.get_model(self.config.llm_config.model_name)
+            if self._model.needs_key:
+                key = get_key(None, key_alias=self._model.needs_key)
+                self._model.key = key
+
+        return self._model
+
+    def initialize_model(self, **kwargs):
+        td = self.training_data
+        s = td.slice
+        if not s[0] and not s[1]:
+            rag_collection = td.collection
+        else:
+            base_collection = td.collection
+            objs = base_collection.find({}, offset=s[0], limit=s[1] - s[0]).rows
+            db = base_collection.parent
+            rag_collection = db.get_collection(f"{base_collection.alias}__rag_{s[0]}_{s[1]}", create_if_not_exists=True)
+            rag_collection.insert(objs)
+        rag_collection.attach_indexer("llm", auto_index=False)
+        self.rag_collection = rag_collection
+
+    def object_to_text(self, object: OBJECT) -> str:
+        return yaml.dump(object)
+
+    def derive(self, object: OBJECT) -> Optional[Inference]:
+        import llm
+        from tiktoken import encoding_for_model
+
+        from linkml_store.utils.llm_utils import get_token_limit, render_formatted_text
+
+        model: llm.Model = self.model
+        model_name = self.config.llm_config.model_name
+        feature_attributes = self.config.feature_attributes
+        target_attributes = self.config.target_attributes
+        num_examples = self.config.llm_config.number_of_few_shot_examples or 5
+        query_text = self.object_to_text(object)
+        if not self.rag_collection.indexers:
+            raise ValueError("RAG collection must have an indexer attached")
+        rs = self.rag_collection.search(query_text, limit=num_examples, index_name="llm")
+        examples = rs.rows
+        if not examples:
+            raise ValueError(f"No examples found for {query_text}; size = {self.rag_collection.size()}")
+        prompt_clauses = []
+        for example in examples:
+            input_obj = {k: example.get(k, None) for k in feature_attributes}
+            output_obj = {k: example.get(k, None) for k in target_attributes}
+            prompt_clause = (
+                "---\nExample:\n"
+                f"## INPUT:\n{self.object_to_text(input_obj)}\n"
+                f"## OUTPUT:\n{self.object_to_text(output_obj)}\n"
+            )
+            prompt_clauses.append(prompt_clause)
+        query_obj = {k: object.get(k, None) for k in feature_attributes}
+        query_text = self.object_to_text(query_obj)
+        prompt_end = "---\nQuery:\n" f"## INPUT:\n{query_text}\n" "## OUTPUT:\n"
+        system_prompt = SYSTEM_PROMPT.format(llm_config=self.config.llm_config)
+
+        def make_text(texts):
+            return "\n".join(prompt_clauses) + prompt_end
+
+        try:
+            encoding = encoding_for_model(model_name)
+        except KeyError:
+            encoding = encoding_for_model("gpt-4")
+        token_limit = get_token_limit(model_name)
+        prompt = render_formatted_text(make_text, prompt_clauses, encoding, token_limit)
+        logger.info(f"Prompt: {prompt}")
+        response = model.prompt(prompt, system_prompt)
+        yaml_str = response.text()
+        logger.info(f"Response: {yaml_str}")
+        try:
+            predicted_object = yaml.safe_load(yaml_str)
+            return Inference(predicted_object=predicted_object)
+        except yaml.parser.ParserError as e:
+            logger.error(f"Error parsing response: {yaml_str}\n{e}")
+            return None

linkml_store/inference/implementations/rule_based_inference_engine.py

@@ -0,0 +1,158 @@
+import logging
+from copy import copy
+from dataclasses import dataclass
+from io import StringIO
+from pathlib import Path
+from typing import Any, ClassVar, Dict, List, Optional, Union
+
+import yaml
+from linkml_map.utils.eval_utils import eval_expr
+from linkml_runtime import SchemaView
+from linkml_runtime.linkml_model.meta import AnonymousClassExpression, ClassRule
+from linkml_runtime.utils.formatutils import underscore
+from pydantic import BaseModel
+
+from linkml_store.api.collection import OBJECT, Collection
+from linkml_store.inference.inference_config import Inference
+from linkml_store.inference.inference_engine import InferenceEngine, ModelSerialization
+
+logger = logging.getLogger(__name__)
+
+
+def expression_matches(ce: AnonymousClassExpression, object: OBJECT) -> bool:
+    """
+    Check if a class expression matches an object.
+
+    :param ce: The class expression
+    :param object: The object
+    :return: True if the class expression matches the object
+    """
+    if ce.any_of:
+        if not any(expression_matches(subce, object) for subce in ce.any_of):
+            return False
+    if ce.all_of:
+        if not all(expression_matches(subce, object) for subce in ce.all_of):
+            return False
+    if ce.none_of:
+        if any(expression_matches(subce, object) for subce in ce.none_of):
+            return False
+    if ce.slot_conditions:
+        for slot in ce.slot_conditions.values():
+            slot_name = slot.name
+            v = object.get(slot_name, None)
+            if slot.equals_string is not None:
+                if slot.equals_string != str(v):
+                    return False
+            if slot.equals_integer is not None:
+                if slot.equals_integer != v:
+                    return False
+            if slot.equals_expression is not None:
+                eval_v = eval_expr(slot.equals_expression, **object)
+                if v != eval_v:
+                    return False
+    return True
+
+
+def apply_rule(rule: ClassRule, object: OBJECT):
+    """
+    Apply a rule to an object.
+
+    Mutates the object
+
+    :param rule: The rule to apply
+    :param object: The object to apply the rule to
+    """
+    for condition in rule.preconditions:
+        if expression_matches(condition, object):
+            for postcondition in rule.postconditions:
+                all_of = [x for x in postcondition.all_of] + [postcondition]
+                for pc in all_of:
+                    sc = pc.slot_condition
+                    if sc:
+                        if sc.equals_string:
+                            object[sc.name] = sc.equals_string
+                        if sc.equals_integer:
+                            object[sc.name] = sc.equals_integer
+                        if sc.equals_expression:
+                            object[sc.name] = eval_expr(sc.equals_expression, **object)
+    return object
+
+
+@dataclass
+class RuleBasedInferenceEngine(InferenceEngine):
+    """
+    TODO
+
+    """
+
+    class_rules: Optional[List[ClassRule]] = None
+    slot_rules: Optional[Dict[str, List[ClassRule]]] = None
+    slot_expressions: Optional[Dict[str, str]] = None
+
+    PERSIST_COLS: ClassVar = ["config", "class_rules", "slot_rules", "slot_expressions"]
+
+    def initialize_model(self, **kwargs):
+        td = self.training_data
+        collection: Collection = td.collection
+        cd = collection.class_definition()
+        sv: SchemaView = collection.parent.schema_view
+        class_rules = cd.rules
+        if class_rules:
+            self.class_rules = class_rules
+        for slot in sv.class_induced_slots(cd.name):
+            if slot.equals_expression:
+                self.slot_expressions[slot.name] = slot.equals_expression
+
+    def derive(self, object: OBJECT) -> Optional[Inference]:
+        object = copy(object)
+        if self.class_rules:
+            for rule in self.class_rules:
+                apply_rule(rule, object)
+        object = {underscore(k): v for k, v in object.items()}
+        if self.slot_expressions:
+            for slot, expr in self.slot_expressions.items():
+                print(f"EVAL {object}")
+                v = eval_expr(expr, **object)
+                if v is not None:
+                    object[slot] = v
+        return Inference(predicted_object=object)
+
+    def import_model_from(self, inference_engine: InferenceEngine, **kwargs):
+        io = StringIO()
+        inference_engine.export_model(io, model_serialization=ModelSerialization.LINKML_EXPRESSION)
+        config = inference_engine.config
+        if len(config.target_attributes) != 1:
+            raise ValueError("Can only import models with a single target attribute")
+        target_attribute = config.target_attributes[0]
+        if self.slot_expressions is None:
+            self.slot_expressions = {}
+        self.slot_expressions[target_attribute] = io.getvalue()
+
+    def save_model(self, output: Union[str, Path]) -> None:
+        """
+        Save the trained model and related data to a file.
+
+        :param output: Path to save the model
+        """
+
+        def _serialize_value(v: Any) -> Any:
+            if isinstance(v, BaseModel):
+                return v.model_dump(exclude_unset=True)
+            return v
+
+        model_data = {k: _serialize_value(getattr(self, k)) for k in self.PERSIST_COLS}
+        with open(output, "w", encoding="utf-8") as f:
+            yaml.dump(model_data, f)
+
+    @classmethod
+    def load_model(cls, file_path: Union[str, Path]) -> "RuleBasedInferenceEngine":
+        model_data = yaml.safe_load(open(file_path))
+
+        engine = cls(config=model_data["config"])
+        for k, v in model_data.items():
+            if k == "config":
+                continue
+            setattr(engine, k, v)
+
+        logger.info(f"Model loaded from {file_path}")
+        return engine

linkml_store/inference/implementations/sklearn_inference_engine.py

@@ -0,0 +1,290 @@
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, ClassVar, Dict, List, Optional, TextIO, Type, Union
+
+import numpy as np
+import pandas as pd
+from sklearn.model_selection import cross_val_score
+from sklearn.preprocessing import LabelEncoder, MultiLabelBinarizer, OneHotEncoder
+from sklearn.tree import DecisionTreeClassifier
+
+from linkml_store.api.collection import OBJECT
+from linkml_store.inference.implementations.rule_based_inference_engine import RuleBasedInferenceEngine
+from linkml_store.inference.inference_config import Inference, InferenceConfig
+from linkml_store.inference.inference_engine import InferenceEngine, ModelSerialization
+from linkml_store.utils.sklearn_utils import tree_to_nested_expression, visualize_decision_tree
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SklearnInferenceEngine(InferenceEngine):
+    config: InferenceConfig
+    classifier: Any = None
+    encoders: Dict[str, Any] = field(default_factory=dict)
+    transformed_features: List[str] = field(default_factory=list)
+    transformed_targets: List[str] = field(default_factory=list)
+    skip_features: List[str] = field(default_factory=list)
+    categorical_encoder_class: Optional[Type[Union[OneHotEncoder, MultiLabelBinarizer]]] = None
+    maximum_proportion_distinct_features: float = 0.2
+    confidence: float = 0.0
+
+    strict: bool = False
+
+    PERSIST_COLS: ClassVar = [
+        "config",
+        "classifier",
+        "encoders",
+        "transformed_features",
+        "transformed_targets",
+        "skip_features",
+        "confidence",
+    ]
+
+    def _get_encoder(self, v: Union[List[Any], Any]) -> Any:
+        if isinstance(v, list):
+            if all(isinstance(x, list) for x in v):
+                return MultiLabelBinarizer()
+            elif all(isinstance(x, str) for x in v):
+                return OneHotEncoder(sparse_output=False, handle_unknown="ignore")
+            elif all(isinstance(x, (int, float)) for x in v):
+                return None
+            else:
+                raise ValueError("Mixed data types in the list are not supported")
+        else:
+            if hasattr(v, "dtype"):
+                if v.dtype == "object" or v.dtype.name == "category":
+                    if isinstance(v.iloc[0], list):
+                        return MultiLabelBinarizer()
+                    elif self.categorical_encoder_class:
+                        return self.categorical_encoder_class(handle_unknown="ignore")
+                    else:
+                        return OneHotEncoder(sparse_output=False, handle_unknown="ignore")
+                elif v.dtype.kind in "biufc":
+                    return None
+        raise ValueError("Unable to determine appropriate encoder for the input data")
+
+    def _is_complex_column(self, column: pd.Series) -> bool:
+        """Check if the column contains complex data types like lists or dicts."""
+        # MV_TYPE = (list, dict)
+        MV_TYPE = (list,)
+        return (column.dtype == "object" or column.dtype == "category") and any(
+            isinstance(x, MV_TYPE) for x in column.dropna()
+        )
+
+    def _get_unique_values(self, column: pd.Series) -> set:
+        """Get unique values from a column, handling list-type data."""
+        if self._is_complex_column(column):
+            # For columns with lists, flatten the lists and get unique values
+            return set(
+                item for sublist in column.dropna() for item in (sublist if isinstance(sublist, list) else [sublist])
+            )
+        else:
+            return set(column.unique())
+
+    def initialize_model(self, **kwargs):
+        logger.info(f"Initializing model with config: {self.config}")
+        df = self.training_data.as_dataframe(flattened=True)
+        logger.info(f"Training data shape: {df.shape}")
+        target_cols = self.config.target_attributes
+        feature_cols = self.config.feature_attributes
+        if len(target_cols) != 1:
+            raise ValueError("Only one target column is supported")
+        if not feature_cols:
+            feature_cols = df.columns.difference(target_cols).tolist()
+            self.config.feature_attributes = feature_cols
+        target_col = target_cols[0]
+        logger.info(f"Feature columns: {feature_cols}")
+        X = df[feature_cols].copy()
+        logger.info(f"Target column: {target_col}")
+        y = df[target_col].copy()
+
+        # find list of features to skip (categorical with > N categories)
+        skip_features = []
+        for col in X.columns:
+            unique_values = self._get_unique_values(X[col])
+            if len(unique_values) > self.maximum_proportion_distinct_features * len(X[col]):
+                skip_features.append(col)
+            if False and (X[col].dtype == "object" or X[col].dtype.name == "category"):
+                if len(X[col].unique()) > self.maximum_proportion_distinct_features * len(X[col]):
+                    skip_features.append(col)
+        self.skip_features = skip_features
+        X = X.drop(skip_features, axis=1)
+        logger.info(f"Skipping features: {skip_features}")
+
+        # Encode features
+        encoded_features = []
+        for col in X.columns:
+            logger.info(f"Checking whether to encode: {col}")
+            col_encoder = self._get_encoder(X[col])
+            if col_encoder:
+                self.encoders[col] = col_encoder
+                if isinstance(col_encoder, OneHotEncoder):
+                    encoded = col_encoder.fit_transform(X[[col]])
+                    feature_names = col_encoder.get_feature_names_out([col])
+                    encoded_df = pd.DataFrame(encoded, columns=feature_names, index=X.index)
+                    X = pd.concat([X.drop(col, axis=1), encoded_df], axis=1)
+                    encoded_features.extend(feature_names)
+                elif isinstance(col_encoder, MultiLabelBinarizer):
+                    encoded = col_encoder.fit_transform(X[col])
+                    feature_names = [f"{col}_{c}" for c in col_encoder.classes_]
+                    encoded_df = pd.DataFrame(encoded, columns=feature_names, index=X.index)
+                    X = pd.concat([X.drop(col, axis=1), encoded_df], axis=1)
+                    encoded_features.extend(feature_names)
+                else:
+                    X[col] = col_encoder.fit_transform(X[col])
+                    encoded_features.append(col)
+            else:
+                encoded_features.append(col)
+
+        self.transformed_features = encoded_features
+        logger.info(f"Encoded features: {self.transformed_features}")
+        logger.info(f"Number of features after encoding: {len(self.transformed_features)}")
+
+        # Encode target
+        # y_encoder = LabelEncoder()
+        y_encoder = self._get_encoder(y)
+        if isinstance(y_encoder, OneHotEncoder):
+            y_encoder = LabelEncoder()
+        # self.encoders[target_col] = y_encoder
+        if y_encoder:
+            self.encoders[target_col] = y_encoder
+            y = y_encoder.fit_transform(y.values.ravel())  # Convert to 1D numpy array
+            self.transformed_targets = y_encoder.classes_
+
+        logger.info(f"Fitting model with features: {X.columns}")
+        clf = DecisionTreeClassifier(random_state=42)
+        clf.fit(X, y)
+        self.classifier = clf
+        logger.info("Model fit complete")
+        cv_scores = cross_val_score(self.classifier, X, y, cv=5)
+        self.confidence = cv_scores.mean()
+        logger.info(f"Cross-validation scores: {cv_scores}")
+
+    def derive(self, object: OBJECT) -> Optional[Inference]:
+        object = self._normalize(object)
+        new_X = pd.DataFrame([object])
+
+        # Apply encodings
+        encoded_features = {}
+        for col in self.config.feature_attributes:
+            if col in self.skip_features:
+                continue
+            if col in self.encoders:
+                encoder = self.encoders[col]
+                if isinstance(encoder, OneHotEncoder):
+                    encoded = encoder.transform(new_X[[col]])
+                    feature_names = encoder.get_feature_names_out([col])
+                    for i, name in enumerate(feature_names):
+                        encoded_features[name] = encoded[0, i]
+                elif isinstance(encoder, MultiLabelBinarizer):
+                    encoded = encoder.transform(new_X[col])
+                    feature_names = [f"{col}_{c}" for c in encoder.classes_]
+                    for i, name in enumerate(feature_names):
+                        encoded_features[name] = encoded[0, i]
+                else:  # LabelEncoder or similar
+                    encoded_features[col] = encoder.transform(new_X[col].astype(str))[0]
+            else:
+                encoded_features[col] = new_X[col].iloc[0]
+
+        # Ensure all expected features are present and in the correct order
+        final_features = []
+        for feature in self.transformed_features:
+            if feature in encoded_features:
+                final_features.append(encoded_features[feature])
+            else:
+                final_features.append(0)  # or some other default value
+
+        # Create the final input array
+        new_X_array = np.array(final_features).reshape(1, -1)
+
+        logger.info(f"Input features: {self.transformed_features}")
+        logger.info(f"Number of input features: {len(self.transformed_features)}")
+
+        predictions = self.classifier.predict(new_X_array)
+        target_attribute = self.config.target_attributes[0]
+        y_encoder = self.encoders.get(target_attribute)
+
+        if y_encoder:
+            v = y_encoder.inverse_transform(predictions)
+        else:
+            v = predictions
+
+        predicted_object = {target_attribute: v[0]}
+        logger.info(f"Predicted object: {predicted_object}")
+        return Inference(predicted_object=predicted_object, confidence=self.confidence)
+
+    def _normalize(self, object: OBJECT) -> OBJECT:
+        return {k: object.get(k, None) for k in self.config.feature_attributes}
+
+    def export_model(
+        self, output: Optional[Union[str, Path, TextIO]], model_serialization: ModelSerialization = None, **kwargs
+    ):
+        def as_file():
+            if isinstance(output, (str, Path)):
+                return open(output, "w")
+            return output
+
+        if model_serialization is None:
+            if isinstance(output, (str, Path)):
+                model_serialization = ModelSerialization.from_filepath(output)
+            if model_serialization is None:
+                model_serialization = ModelSerialization.JOBLIB
+
+        if model_serialization == ModelSerialization.LINKML_EXPRESSION:
+            expr = tree_to_nested_expression(
+                self.classifier,
+                self.transformed_features,
+                self.encoders.keys(),
+                feature_encoders=self.encoders,
+                target_encoder=self.encoders.get(self.config.target_attributes[0]),
+            )
+            as_file().write(expr)
+        elif model_serialization == ModelSerialization.JOBLIB:
+            self.save_model(output)
+        elif model_serialization == ModelSerialization.RULE_BASED:
+            rbie = RuleBasedInferenceEngine(config=self.config)
+            rbie.import_model_from(self)
+            rbie.save_model(output)
+        elif model_serialization == ModelSerialization.PNG:
+            visualize_decision_tree(self.classifier, self.transformed_features, self.transformed_targets, output)
+        else:
+            raise ValueError(f"Unsupported model serialization: {model_serialization}")
+
+    def save_model(self, output: Union[str, Path]) -> None:
+        """
+        Save the trained model and related data to a file.
+
+        :param output: Path to save the model
+        """
+        import joblib
+
+        if self.classifier is None:
+            raise ValueError("Model has not been trained. Call initialize_model() first.")
+
+        # Use self.PERSIST_COLS
+        model_data = {k: getattr(self, k) for k in self.PERSIST_COLS}
+
+        joblib.dump(model_data, output)
+
+    @classmethod
+    def load_model(cls, file_path: Union[str, Path]) -> "SklearnInferenceEngine":
+        """
+        Load a trained model and related data from a file.
+
+        :param file_path: Path to the saved model
+        :return: SklearnInferenceEngine instance with loaded model
+        """
+        import joblib
+
+        model_data = joblib.load(file_path)
+
+        engine = cls(config=model_data["config"])
+        for k, v in model_data.items():
+            if k == "config":
+                continue
+            setattr(engine, k, v)
+
+        logger.info(f"Model loaded from {file_path}")
+        return engine

linkml_store/inference/inference_config.py

@@ -0,0 +1,62 @@
+import logging
+from typing import List, Optional, Tuple
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from linkml_store.api.collection import OBJECT
+from linkml_store.utils.format_utils import Format, load_objects
+
+logger = logging.getLogger(__name__)
+
+
+class LLMConfig(BaseModel, extra="forbid"):
+    """
+    Configuration for the LLM indexer.
+    """
+
+    model_config = ConfigDict(protected_namespaces=())
+
+    model_name: str = "gpt-4o-mini"
+    token_limit: Optional[int] = None
+    number_of_few_shot_examples: Optional[int] = None
+    role: str = "Domain Expert"
+    cached_embeddings_database: Optional[str] = None
+    cached_embeddings_collection: Optional[str] = None
+    text_template: Optional[str] = None
+    text_template_syntax: Optional[str] = None
+
+
+class InferenceConfig(BaseModel, extra="forbid"):
+    """
+    Configuration for inference engines.
+    """
+
+    target_attributes: Optional[List[str]] = None
+    feature_attributes: Optional[List[str]] = None
+    train_test_split: Optional[Tuple[float, float]] = None
+    llm_config: Optional[LLMConfig] = None
+
+    @classmethod
+    def from_file(cls, file_path: str, format: Optional[Format] = None) -> "InferenceConfig":
+        """
+        Load an inference config from a file.
+
+        :param file_path: Path to the file.
+        :param format: Format of the file (YAML is recommended).
+        :return: InferenceConfig
+        """
+        if format and format.is_xsv():
+            logger.warning("XSV format is not recommended for inference config files")
+        objs = load_objects(file_path, format=format)
+        if len(objs) != 1:
+            raise ValueError(f"Expected 1 object, got {len(objs)}")
+        return cls(**objs[0])
+
+
+class Inference(BaseModel, extra="forbid"):
+    """
+    Result of an inference derivation.
+    """
+
+    predicted_object: OBJECT = Field(..., description="The predicted object.")
+    confidence: Optional[float] = Field(default=None, description="The confidence of the prediction.", le=1.0, ge=0.0)