ibm-watsonx-orchestrate-evaluation-framework 1.0.3__py3-none-any.whl → 1.1.8b0__py3-none-any.whl

wxo_agentic_evaluation/referenceless_eval/metrics/field.py

@@ -0,0 +1,266 @@
+from __future__ import annotations
+
+from abc import ABC
+from typing import Any, Dict, List, Literal, Optional, Type, TypeVar
+
+from pydantic import BaseModel
+from pydantic import Field as PydanticField
+from pydantic import PrivateAttr, model_validator
+
+JSONType = Literal["integer", "number", "string", "boolean", "object", "array"]
+TField = TypeVar("TField", bound="BaseField")
+BaseFieldRegistry: List[Type[BaseField]] = []
+
+
+class BaseField(BaseModel, ABC):
+    """
+    Abstract representation of a single metric field.
+
+    Attributes:
+        name: Identifier of the field (used as JSON key).
+        json_type: JSON Schema type of the field.
+        description: Human-friendly description of the field's purpose.
+        jsonschema_extra: Additional JSONSchema keywords (e.g., enum, pattern).
+        extra_params: Non-JSONSchema attributes (e.g., thresholds).
+    """
+
+    name: str
+    json_type: JSONType
+    description: str = PydanticField(
+        "No description provided. Please specify what this field represents.",
+        description="A clear description of this field's meaning.",
+    )
+    jsonschema_extra: Dict[str, Any] = PydanticField(
+        default_factory=dict,
+        description="Additional JSONSchema constraints for this field.",
+    )
+    extra_params: Dict[str, Any] = PydanticField(
+        default_factory=dict,
+        description="Extra parameters not included in the JSONSchema (e.g., thresholds).",
+    )
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        if not getattr(cls, "__abstract__", False):
+            BaseFieldRegistry.insert(0, cls)
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        """Override in subclasses to signal compatibility with a JSONSchema snippet."""
+        return False
+
+    @classmethod
+    def from_jsonschema(cls, name: str, schema: Dict[str, Any]) -> BaseField:
+        """
+        Instantiate the appropriate Field subclass from a JSONSchema property.
+        The first subclass whose `can_handle` returns True is used.
+        Falls back to GenericField.
+        """
+        for field_cls in BaseFieldRegistry:
+            if field_cls.can_handle(name, schema):
+                desc = schema.get("description", "")
+                extra = {
+                    k: v
+                    for k, v in schema.items()
+                    if k not in ("type", "description")
+                }
+                return field_cls(
+                    name=name,
+                    json_type=schema.get("type", "string"),
+                    description=desc,
+                    jsonschema_extra=extra,
+                    extra_params={},
+                )
+        return GenericField(
+            name=name,
+            json_type=schema.get("type", "string"),
+            description=schema.get("description", ""),
+            jsonschema_extra={
+                k: v
+                for k, v in schema.items()
+                if k not in ("type", "description")
+            },
+            extra_params={},
+        )
+
+    def to_jsonschema(self) -> Dict[str, Any]:
+        return {
+            "type": self.json_type,
+            "description": self.description,
+            **self.jsonschema_extra,
+        }
+
+    # --- Getters and Setters ---
+
+    def get_name(self) -> str:
+        return self.name
+
+    def set_name(self, name: str) -> None:
+        self.name = name
+
+    def get_description(self) -> str:
+        return self.description
+
+    def set_description(self, description: str) -> None:
+        self.description = description
+
+    def get_jsonschema_extra(self) -> Dict[str, Any]:
+        return dict(self.jsonschema_extra)
+
+    def set_jsonschema_extra(self, extra: Dict[str, Any]) -> None:
+        self.jsonschema_extra = extra
+
+    def get_extra_param(self, key: str) -> Any:
+        return self.extra_params.get(key)
+
+    def set_extra_param(self, key: str, value: Any) -> None:
+        self.extra_params[key] = value
+
+
+class NumericField(BaseField):
+    """
+    Numeric field (integer or number) with optional thresholds.
+    The `extra_params` dict may include:
+      - threshold_low: minimal acceptable value (for validation)
+      - threshold_high: maximal acceptable value
+    """
+
+    threshold_low: Optional[float] = PydanticField(
+        None,
+        description="Lower bound for correctness checks (not in JSONSchema).",
+    )
+    threshold_high: Optional[float] = PydanticField(
+        None,
+        description="Upper bound for correctness checks (not in JSONSchema).",
+    )
+
+    __abstract__ = False
+
+    @model_validator(mode="before")
+    def extract_thresholds(cls, values: Dict[str, Any]) -> Dict[str, Any]:
+        extra = values.get("jsonschema_extra", {})
+        if "threshold_low" in extra:
+            values["threshold_low"] = extra["threshold_low"]
+        if "threshold_high" in extra:
+            values["threshold_high"] = extra["threshold_high"]
+        return values
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return schema.get("type") in ("integer", "number")
+
+    @classmethod
+    def from_jsonschema(cls, name: str, schema: Dict[str, Any]) -> NumericField:
+        """
+        Create a NumericField from a JSONSchema property.
+        """
+        return NumericField(
+            name=name,
+            json_type=schema.get("type", "number"),
+            description=schema.get("description", ""),
+            jsonschema_extra={
+                k: v
+                for k, v in schema.items()
+                if k not in ("type", "description")
+            },
+            extra_params={},
+        )
+
+    def to_jsonschema(self) -> Dict[str, Any]:
+        return super().to_jsonschema()
+
+    def is_within_threshold(self, value: float) -> bool:
+        if self.threshold_low is not None and value < self.threshold_low:
+            return False
+        if self.threshold_high is not None and value > self.threshold_high:
+            return False
+        return True
+
+
+class EnumField(BaseField):
+    """
+    Field whose value must be one of a fixed set of options.
+    Expects `jsonschema_extra["enum"]` to be a list of allowed values.
+    """
+
+    __abstract__ = False
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return "enum" in schema
+
+
+class ExplanationField(BaseField):
+    """
+    Free-form explanation of the metric's reasoning.
+    """
+
+    __abstract__ = False
+
+    def __init__(self, **data: Any):
+        data.setdefault(
+            "description",
+            "A detailed, step-by-step explanation of the reasoning behind the metric's value.",
+        )
+        super().__init__(**data)
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return name.lower() == "explanation" and schema.get("type") == "string"
+
+
+class EvidenceField(BaseField):
+    """
+    The specific quote or reference that supports the metric's evaluation.
+    """
+
+    __abstract__ = False
+
+    def __init__(self, **data: Any):
+        data.setdefault(
+            "description",
+            "The exact quote or reference from the input or context that justifies the metric's value.",
+        )
+        super().__init__(**data)
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return name.lower() == "evidence" and schema.get("type") == "string"
+
+
+class CorrectionField(BaseField):
+    """
+    A structured suggestion (as JSON) for correcting or improving the output.
+    """
+
+    __abstract__ = False
+
+    def __init__(self, **data: Any):
+        data.setdefault(
+            "description",
+            "A JSON-formatted suggestion for how to correct or improve the output if needed.",
+        )
+        super().__init__(**data)
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return name.lower() == "correction" and schema.get("type") == "object"
+
+
+class GenericField(BaseField):
+    """
+    Fallback field type for any property not handled by other classes.
+    """
+
+    __abstract__ = False
+
+    def __init__(self, **data: Any):
+        data.setdefault(
+            "description",
+            f"A generic field named '{data.get('name')}' of type {data.get('json_type')}.",
+        )
+        super().__init__(**data)
+
+    @classmethod
+    def can_handle(cls, name: str, schema: Dict[str, Any]) -> bool:
+        return True

wxo_agentic_evaluation/referenceless_eval/metrics/metric.py

@@ -0,0 +1,344 @@
+from __future__ import annotations
+
+import json
+from typing import Any, Dict, List, Optional, Set, Tuple, Type, TypeVar
+
+from wxo_agentic_evaluation.referenceless_eval.metrics.field import (
+    BaseField,
+    CorrectionField,
+    EvidenceField,
+    ExplanationField,
+    NumericField,
+)
+
+TMetric = TypeVar("TMetric", bound="Metric")
+
+
+class Metric:
+    """
+    Abstract representation of an evaluation metric composed of multiple fields.
+    """
+
+    def __init__(
+        self,
+        name: str,
+        description: str,
+        fields: Optional[List[BaseField]] = None,
+        required: Optional[List[str]] = None,
+        additional_properties: bool = True,
+    ) -> None:
+        """
+        Args:
+            name: Unique metric identifier.
+            description: Full description of what this metric measures.
+            fields: List of BaseField instances composing this metric.
+            required: List of field names that must appear in results.
+                      Defaults to all provided fields.
+        """
+        self.name = name
+        self.description = description
+        self.fields: List[BaseField] = fields or []
+        self.additional_properties = additional_properties
+        # Determine required fields
+        if required is not None:
+            self.required_fields: Set[str] = set(required)
+        else:
+            self.required_fields: Set[str] = {f.name for f in self.fields}
+
+        # Validate required_fields
+        known = {f.name for f in self.fields}
+        missing = self.required_fields - known
+        if missing:
+            raise ValueError(
+                f"Required fields {missing} not among metric fields {known}"
+            )
+
+    def to_jsonschema(self) -> Dict[str, Any]:
+        """
+        Build a JSONSchema representation of this metric.
+
+        Returns:
+            A dict with keys:
+              - title: self.name
+              - description: self.description
+              - type: "object"
+              - properties: mapping field.name → field.to_jsonschema()
+              - required: list of required field names
+        """
+        props: Dict[str, Any] = {f.name: f.to_jsonschema() for f in self.fields}
+        return {
+            "title": self.name,
+            "description": self.description,
+            "type": "object",
+            "properties": props,
+            "required": sorted(self.required_fields),
+            "additionalProperties": self.additional_properties,
+        }
+
+    def add_field(self, field: BaseField, required: bool = True) -> None:
+        """
+        Add a new field to this metric.
+
+        Args:
+            field: BaseField instance.
+            required: Whether this field must appear in results.
+        """
+        if any(f.name == field.name for f in self.fields):
+            raise ValueError(f"Field '{field.name}' already defined")
+        self.fields.append(field)
+        if required:
+            self.required_fields.add(field.name)
+
+    def remove_field(self, name: str) -> None:
+        """
+        Remove a field by name.
+
+        Args:
+            name: Name of field to remove.
+        """
+        self.fields = [f for f in self.fields if f.name != name]
+        self.required_fields.discard(name)
+
+    @classmethod
+    def from_jsonschema(cls: Type[TMetric], schema: Dict[str, Any]) -> Metric:
+        """
+        Reconstruct a Metric from a JSONSchema dict.
+
+        Args:
+            schema: dict with 'title', 'description', 'properties', 'required'.
+
+        Returns:
+            Metric instance with fields populated.
+        """
+        name: str = schema.get("title", "")
+        description: str = schema.get("description", "")
+        props: Dict[str, Any] = schema.get("properties", {})
+        required: List[str] = schema.get("required", [])
+        additional_props: bool = schema.get("additionalProperties", True)
+        fields: List[BaseField] = []
+        for fname, fschema in props.items():
+            # If type is number or integer, use NumericField
+            if fschema.get("type") in ("number", "integer"):
+                field = NumericField.from_jsonschema(fname, fschema)
+            else:
+                field = BaseField.from_jsonschema(fname, fschema)
+            fields.append(field)
+        return cls(
+            name=name,
+            description=description,
+            fields=fields,
+            required=required,
+            additional_properties=additional_props,
+        )
+
+    def is_important(
+        self, result: Dict[str, Any]
+    ) -> Tuple[bool, Optional[str]]:
+        """
+        A result is 'important' if its confidence lies within the defined confidence thresholds.
+
+        Args:
+            result: Parsed metric result with at least 'confidence'.
+
+        Returns:
+            (important: bool, reason: Optional[str])
+        """
+        try:
+            conf = float(result.get("confidence", 0.0))
+        except (TypeError, ValueError):
+            return False, "Invalid confidence value"
+        # locate the confidence field
+        conf_field = next(
+            (f for f in self.fields if f.name == "confidence"), None
+        )
+        if isinstance(conf_field, NumericField):
+            ok = conf_field.is_within_threshold(conf)
+            reason = (
+                None
+                if ok
+                else f"Confidence {conf} outside [{conf_field.threshold_low},{conf_field.threshold_high}]"
+            )
+            return ok, reason
+        return False, "Confidence field not defined"
+
+    def is_correct(self, result: Dict[str, Any]) -> Tuple[bool, Optional[str]]:
+        """
+        A result is 'correct' if it is important AND its output lies within thresholds.
+
+        Args:
+            result: Parsed metric result with 'output' and 'confidence'.
+
+        Returns:
+            (correct: bool, reason: Optional[str])
+        """
+        important, imp_reason = self.is_important(result)
+        if not important:
+            return True, f"Not important: {imp_reason}"
+        # check output
+        try:
+            val = float(result.get("output", 0.0))
+        except (TypeError, ValueError):
+            return False, "Invalid output value"
+        out_field = next((f for f in self.fields if f.name == "output"), None)
+        if isinstance(out_field, NumericField):
+            ok = out_field.is_within_threshold(val)
+            reason = (
+                None
+                if ok
+                else f"Output {val} outside [{out_field.threshold_low},{out_field.threshold_high}]"
+            )
+            return ok, reason
+        return False, "Output field not defined"
+
+    def parse_response(self, response: str) -> Dict[str, Any]:
+        """
+        Parse a raw response string into a structured dict.
+
+        Args:
+            response: Raw response string.
+
+        Returns:
+            Parsed response as a dict.
+        """
+        # Default implementation: assume JSON string
+        try:
+            return json.loads(response)
+        except json.JSONDecodeError as e:
+            raise ValueError(f"Failed to parse response: {e}") from e
+
+
+class StandardMetric(Metric):
+    """
+    A standard metric with common fields:
+      - explanation: string, detailed reasoning.
+      - evidence: string, supporting quote or reference.
+      - output: numeric value within specified range.
+      - confidence: numeric confidence within specified range.
+      - correction: object, structured suggestion for improvement.
+    Also provides convenience methods `is_important` and `is_correct`.
+    """
+
+    def __init__(
+        self,
+        name: str,
+        description: str,
+        *,
+        output_range: Tuple[float, float] = (0.0, 1.0),
+        confidence_range: Tuple[float, float] = (0.0, 1.0),
+    ) -> None:
+        """
+        Args:
+            name: Metric identifier.
+            description: Explanation of what the metric measures.
+            output_range: (min, max) allowed for the 'output' field.
+            confidence_range: (min, max) for the 'confidence' field.
+
+        Fields created:
+          - explanation: "A detailed, step-by-step explanation of the reasoning."
+          - evidence: "The exact quote or evidence supporting the reasoning."
+          - output: numeric in output_range
+          - confidence: numeric in confidence_range
+          - correction: structured suggestion if output below threshold
+        """
+        # Prepare fields
+        min_out, max_out = output_range
+        min_conf, max_conf = confidence_range
+
+        explanation = ExplanationField(
+            name="explanation",
+            json_type="string",
+            description="A detailed, step-by-step explanation of the reasoning behind the output value.",
+        )
+        evidence = EvidenceField(
+            name="evidence",
+            json_type="string",
+            description="The exact quote or reference that supports the output value.",
+        )
+        output = NumericField(
+            name="output",
+            json_type=(
+                "number"
+                if isinstance(min_out, float) or isinstance(max_out, float)
+                else "integer"
+            ),
+            description=f"Primary numeric score for this metric (range {min_out} to {max_out}).",
+            jsonschema_extra={"minimum": min_out, "maximum": max_out},
+            extra_params={"threshold_low": min_out, "threshold_high": max_out},
+        )
+        confidence = NumericField(
+            name="confidence",
+            json_type="number",
+            description=f"Confidence in the output value (range {min_conf} to {max_conf}).",
+            jsonschema_extra={"minimum": min_conf, "maximum": max_conf},
+            extra_params={
+                "threshold_low": min_conf,
+                "threshold_high": max_conf,
+            },
+        )
+        correction = CorrectionField(
+            name="correction",
+            json_type="object",
+            description="Structured suggestion for how to correct or improve the output if needed.",
+        )
+
+        fields = [explanation, evidence, output, confidence, correction]
+        super().__init__(name=name, description=description, fields=fields)
+
+    def is_important(
+        self, result: Dict[str, Any]
+    ) -> Tuple[bool, Optional[str]]:
+        """
+        A result is 'important' if its confidence lies within the defined confidence thresholds.
+
+        Args:
+            result: Parsed metric result with at least 'confidence'.
+
+        Returns:
+            (important: bool, reason: Optional[str])
+        """
+        try:
+            conf = float(result.get("confidence", 0.0))
+        except (TypeError, ValueError):
+            return False, "Invalid confidence value"
+        # locate the confidence field
+        conf_field = next(
+            (f for f in self.fields if f.name == "confidence"), None
+        )
+        if isinstance(conf_field, NumericField):
+            ok = conf_field.is_within_threshold(conf)
+            reason = (
+                None
+                if ok
+                else f"Confidence {conf} outside [{conf_field.threshold_low},{conf_field.threshold_high}]"
+            )
+            return ok, reason
+        return False, "Confidence field not defined"
+
+    def is_correct(self, result: Dict[str, Any]) -> Tuple[bool, Optional[str]]:
+        """
+        A result is 'correct' if it is important AND its output lies within thresholds.
+
+        Args:
+            result: Parsed metric result with 'output' and 'confidence'.
+
+        Returns:
+            (correct: bool, reason: Optional[str])
+        """
+        important, imp_reason = self.is_important(result)
+        if not important:
+            return True, f"Not important: {imp_reason}"
+        # check output
+        try:
+            val = float(result.get("output", 0.0))
+        except (TypeError, ValueError):
+            return False, "Invalid output value"
+        out_field = next((f for f in self.fields if f.name == "output"), None)
+        if isinstance(out_field, NumericField):
+            ok = out_field.is_within_threshold(val)
+            reason = (
+                None
+                if ok
+                else f"Output {val} outside [{out_field.threshold_low},{out_field.threshold_high}]"
+            )
+            return ok, reason
+        return False, "Output field not defined"