arize-phoenix 4.4.4rc4__py3-none-any.whl → 4.4.4rc6__py3-none-any.whl

phoenix/experiments/evaluators/base.py

@@ -0,0 +1,153 @@
+import functools
+import inspect
+from abc import ABC
+from types import MappingProxyType
+from typing import Any, Awaitable, Callable, Optional, Union
+
+from typing_extensions import TypeAlias
+
+from phoenix.experiments.evaluators.utils import validate_signature
+from phoenix.experiments.types import (
+    AnnotatorKind,
+    EvaluationResult,
+    EvaluatorKind,
+    EvaluatorName,
+    EvaluatorOutput,
+    ExampleInput,
+    ExampleMetadata,
+    ExampleOutput,
+    TaskOutput,
+)
+
+
+class Evaluator(ABC):
+    """
+    A helper super class to guide the implementation of an `Evaluator` object.
+    Subclasses must implement either the `evaluate` or `async_evaluate` method.
+    Implementing both methods is recommended, but not required.
+
+    This Class is intended to be subclassed, and should not be instantiated directly.
+    """
+
+    _kind: AnnotatorKind
+    _name: EvaluatorName
+
+    @functools.cached_property
+    def name(self) -> EvaluatorName:
+        if hasattr(self, "_name"):
+            return self._name
+        return self.__class__.__name__
+
+    @functools.cached_property
+    def kind(self) -> EvaluatorKind:
+        if hasattr(self, "_kind"):
+            return self._kind.value
+        return AnnotatorKind.CODE.value
+
+    def __new__(cls, *args: Any, **kwargs: Any) -> "Evaluator":
+        if cls is Evaluator:
+            raise TypeError(f"{cls.__name__} is an abstract class and should not be instantiated.")
+        return object.__new__(cls)
+
+    def evaluate(
+        self,
+        *,
+        output: Optional[TaskOutput] = None,
+        expected: Optional[ExampleOutput] = None,
+        metadata: ExampleMetadata = MappingProxyType({}),
+        input: ExampleInput = MappingProxyType({}),
+        **kwargs: Any,
+    ) -> EvaluationResult:
+        # For subclassing, one should implement either this sync method or the
+        # async version. Implementing both is recommended but not required.
+        raise NotImplementedError
+
+    async def async_evaluate(
+        self,
+        *,
+        output: Optional[TaskOutput] = None,
+        expected: Optional[ExampleOutput] = None,
+        metadata: ExampleMetadata = MappingProxyType({}),
+        input: ExampleInput = MappingProxyType({}),
+        **kwargs: Any,
+    ) -> EvaluationResult:
+        # For subclassing, one should implement either this async method or the
+        # sync version. Implementing both is recommended but not required.
+        return self.evaluate(
+            output=output,
+            expected=expected,
+            metadata=metadata,
+            input=input,
+            **kwargs,
+        )
+
+    def __init_subclass__(cls, is_abstract: bool = False, **kwargs: Any) -> None:
+        super().__init_subclass__(**kwargs)
+        if is_abstract:
+            return
+        evaluate_fn_signature = inspect.signature(Evaluator.evaluate)
+        for super_cls in inspect.getmro(cls):
+            if super_cls in (LLMEvaluator, Evaluator):
+                break
+            if evaluate := super_cls.__dict__.get(Evaluator.evaluate.__name__):
+                assert callable(evaluate), "`evaluate()` method should be callable"
+                # need to remove the first param, i.e. `self`
+                _validate_sig(functools.partial(evaluate, None), "evaluate")
+                return
+            if async_evaluate := super_cls.__dict__.get(Evaluator.async_evaluate.__name__):
+                assert callable(async_evaluate), "`async_evaluate()` method should be callable"
+                # need to remove the first param, i.e. `self`
+                _validate_sig(functools.partial(async_evaluate, None), "async_evaluate")
+                return
+        raise ValueError(
+            f"Evaluator must implement either "
+            f"`def evaluate{evaluate_fn_signature}` or "
+            f"`async def async_evaluate{evaluate_fn_signature}`"
+        )
+
+
+def _validate_sig(fn: Callable[..., Any], fn_name: str) -> None:
+    sig = inspect.signature(fn)
+    validate_signature(sig)
+    for param in sig.parameters.values():
+        if param.kind is inspect.Parameter.VAR_KEYWORD:
+            return
+    else:
+        raise ValueError(f"`{fn_name}` should allow variadic keyword arguments `**kwargs`")
+
+
+class CodeEvaluator(Evaluator, ABC, is_abstract=True):
+    """
+    A convenience super class for defining code evaluators.
+
+    This class is intended to be subclassed, and should not be instantiated directly.
+    """
+
+    _kind = AnnotatorKind.CODE
+
+    def __new__(cls, *args: Any, **kwargs: Any) -> "CodeEvaluator":
+        if cls is CodeEvaluator:
+            raise TypeError(f"{cls.__name__} is an abstract class and should not be instantiated.")
+        return object.__new__(cls)
+
+
+class LLMEvaluator(Evaluator, ABC, is_abstract=True):
+    """
+    A convenience super class for defining LLM evaluators.
+
+    This class is intended to be subclassed, and should not be instantiated directly.
+    """
+
+    _kind = AnnotatorKind.LLM
+
+    def __new__(cls, *args: Any, **kwargs: Any) -> "LLMEvaluator":
+        if cls is LLMEvaluator:
+            raise TypeError(f"{cls.__name__} is an abstract class and should not be instantiated.")
+        return object.__new__(cls)
+
+
+ExperimentEvaluator: TypeAlias = Union[
+    Evaluator,
+    Callable[..., EvaluatorOutput],
+    Callable[..., Awaitable[EvaluatorOutput]],
+]

phoenix/{datasets → experiments}/evaluators/code_evaluators.py

@@ -2,19 +2,14 @@ from __future__ import annotations
 
 import json
 import re
-from typing import TYPE_CHECKING, List, Optional, Union
+from typing import Any, List, Optional, Union
 
-from phoenix.datasets.evaluators._utils import _unwrap_json
-from phoenix.datasets.types import EvaluationResult, Example, ExperimentEvaluator, ExperimentRun
+from phoenix.experiments.evaluators.base import CodeEvaluator
+from phoenix.experiments.types import EvaluationResult, TaskOutput
 
 
-class JSONParsable:
-    annotator_kind = "CODE"
-    name = "JSONParsable"
-
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        assert exp_run.output is not None
-        output = _unwrap_json(exp_run.output.result)
+class JSONParsable(CodeEvaluator):
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
         assert isinstance(output, str), "Experiment run output must be a string"
         try:
             json.loads(output)
@@ -26,18 +21,14 @@ class JSONParsable:
         )
 
 
-class ContainsKeyword:
-    annotator_kind = "CODE"
-
+class ContainsKeyword(CodeEvaluator):
     def __init__(self, keyword: str, name: Optional[str] = None) -> None:
         self.keyword = keyword
-        self.name = name or f"Contains({repr(keyword)})"
+        self._name = name or f"Contains({repr(keyword)})"
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        assert exp_run.output is not None
-        result = _unwrap_json(exp_run.output.result)
-        assert isinstance(result, str), "Experiment run output must be a string"
-        found = self.keyword in result
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
+        assert isinstance(output, str), "Experiment run output must be a string"
+        found = self.keyword in output
         return EvaluationResult(
             score=float(found),
             explanation=(
@@ -47,18 +38,14 @@ class ContainsKeyword:
         )
 
 
-class ContainsAnyKeyword:
-    annotator_kind = "CODE"
-
+class ContainsAnyKeyword(CodeEvaluator):
     def __init__(self, keywords: List[str], name: Optional[str] = None) -> None:
         self.keywords = keywords
-        self.name = name or f"ContainsAny({keywords})"
+        self._name = name or f"ContainsAny({keywords})"
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        assert exp_run.output is not None
-        result = _unwrap_json(exp_run.output.result)
-        assert isinstance(result, str), "Experiment run output must be a string"
-        found = [keyword for keyword in self.keywords if keyword in result]
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
+        assert isinstance(output, str), "Experiment run output must be a string"
+        found = [keyword for keyword in self.keywords if keyword in output]
         if found:
             explanation = f"the keywords {found} were found in the output"
         else:
@@ -69,18 +56,14 @@ class ContainsAnyKeyword:
         )
 
 
-class ContainsAllKeywords:
-    annotator_kind = "CODE"
-
+class ContainsAllKeywords(CodeEvaluator):
     def __init__(self, keywords: List[str], name: Optional[str] = None) -> None:
         self.keywords = keywords
-        self.name = name or f"ContainsAll({keywords})"
+        self._name = name or f"ContainsAll({keywords})"
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        assert exp_run.output is not None
-        result = _unwrap_json(exp_run.output.result)
-        assert isinstance(result, str), "Experiment run output must be a string"
-        not_found = [keyword for keyword in self.keywords if keyword not in result]
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
+        assert isinstance(output, str), "Experiment run output must be a string"
+        not_found = [keyword for keyword in self.keywords if keyword not in output]
         if not_found:
             contains_all = False
             explanation = f"the keywords {not_found} were not found in the output"
@@ -93,21 +76,17 @@ class ContainsAllKeywords:
         )
 
 
-class MatchesRegex:
-    annotator_kind = "CODE"
-
+class MatchesRegex(CodeEvaluator):
     def __init__(self, pattern: Union[str, re.Pattern[str]], name: Optional[str] = None) -> None:
         if isinstance(pattern, str):
             pattern = re.compile(pattern)
         self.pattern = pattern
         assert isinstance(pattern, re.Pattern)
-        self.name = name or f"matches_({pattern})"
+        self._name = name or f"matches_({pattern})"
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        assert exp_run.output is not None
-        result = _unwrap_json(exp_run.output.result)
-        assert isinstance(result, str), "Experiment run output must be a string"
-        matches = self.pattern.findall(result)
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
+        assert isinstance(output, str), "Experiment run output must be a string"
+        matches = self.pattern.findall(output)
         if matches:
             explanation = (
                 f"the substrings {matches} matched the regex pattern {self.pattern.pattern}"
@@ -118,10 +97,3 @@ class MatchesRegex:
             score=float(bool(matches)),
             explanation=explanation,
         )
-
-
-# Someday we'll do typing checking in unit tests.
-if TYPE_CHECKING:
-    _: ExperimentEvaluator
-    _ = JSONParsable()
-    _ = ContainsKeyword("test")

phoenix/{datasets → experiments}/evaluators/llm_evaluators.py

@@ -1,14 +1,23 @@
 import re
-from typing import Callable, Optional, Type
+from types import MappingProxyType
+from typing import Any, Callable, Optional, Type
 
-from phoenix.datasets.evaluators._utils import _unwrap_json
-from phoenix.datasets.types import EvaluationResult, Example, ExperimentEvaluator, ExperimentRun
 from phoenix.evals.models.base import BaseModel as LLMBaseModel
 from phoenix.evals.utils import snap_to_rail
+from phoenix.experiments.evaluators.base import (
+    ExperimentEvaluator,
+    LLMEvaluator,
+)
+from phoenix.experiments.evaluators.utils import unwrap_json
+from phoenix.experiments.types import (
+    EvaluationResult,
+    ExampleInput,
+    ExampleMetadata,
+    TaskOutput,
+)
 
 
-class LLMCriteriaEvaluator:
-    annotator_kind = "LLM"
+class LLMCriteriaEvaluator(LLMEvaluator):
     _base_template = (
         "Determine if the following text is {criteria}. {description}"
         "First, explain step-by-step why you think the text is or is not {criteria}. Then provide "
@@ -37,21 +46,23 @@ class LLMCriteriaEvaluator:
         self.criteria = criteria
         self.description = description
         self.template = self._format_base_template(self.criteria, self.description)
-        self.name = name
+        self._name = name
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        formatted_template = self._format_eval_template(exp_run)
+    def evaluate(self, *, output: Optional[TaskOutput] = None, **_: Any) -> EvaluationResult:
+        formatted_template = self._format_eval_template(output)
         unparsed_response = self.model._generate(formatted_template)
         return self._parse_eval_output(unparsed_response)
 
-    async def async_evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        formatted_template = self._format_eval_template(exp_run)
+    async def async_evaluate(
+        self, *, output: Optional[TaskOutput] = None, **_: Any
+    ) -> EvaluationResult:
+        formatted_template = self._format_eval_template(output)
         unparsed_response = await self.model._async_generate(formatted_template)
         return self._parse_eval_output(unparsed_response)
 
-    def _format_eval_template(self, experiment_run: ExperimentRun) -> str:
-        assert experiment_run.output is not None
-        result = _unwrap_json(experiment_run.output.result)
+    def _format_eval_template(self, output: TaskOutput) -> str:
+        assert output is not None
+        result = unwrap_json(output)
         return self.template.format(text=str(result))
 
     def _parse_eval_output(self, unparsed_response: str) -> EvaluationResult:
@@ -137,8 +148,7 @@ def _parse_label_from_explanation(raw_string: str) -> str:
     return raw_string
 
 
-class RelevanceEvaluator:
-    annotator_kind = "LLM"
+class RelevanceEvaluator(LLMEvaluator):
     template = (
         "Determine if the following response is relevant to the query. In this context, "
         "'relevance' means that the response directly addresses the core question or topic of the "
@@ -162,19 +172,24 @@ class RelevanceEvaluator:
     def __init__(
         self,
         model: LLMBaseModel,
-        get_query: Optional[Callable[[Example, ExperimentRun], str]] = None,
-        get_response: Optional[Callable[[Example, ExperimentRun], str]] = None,
+        get_query: Optional[Callable[[ExampleInput, ExampleMetadata], str]] = None,
+        get_response: Optional[Callable[[Optional[TaskOutput], ExampleMetadata], str]] = None,
         name: str = "RelevanceEvaluator",
     ):
         self.model = model
-        self.name = name
+        self._name = name
         self.get_query = get_query or self._default_get_query
         self.get_response = get_response or self._default_get_response
 
-    def _format_eval_template(self, example: Example, experiment_run: ExperimentRun) -> str:
-        assert experiment_run.output is not None
-        query = self.get_query(example, experiment_run)
-        response = self.get_response(example, experiment_run)
+    def _format_eval_template(
+        self,
+        output: Optional[TaskOutput] = None,
+        input: ExampleInput = MappingProxyType({}),
+        metadata: ExampleMetadata = MappingProxyType({}),
+    ) -> str:
+        assert output is not None
+        query = self.get_query(input, metadata)
+        response = self.get_response(output, metadata)
         return self.template.format(query=query, response=response)
 
     def _parse_eval_output(self, unparsed_response: str) -> EvaluationResult:
@@ -195,19 +210,35 @@ class RelevanceEvaluator:
             metadata={},
         )
 
-    def _default_get_query(self, example: Example, experiment_run: ExperimentRun) -> str:
-        return str(example.input)
+    def _default_get_query(self, input: ExampleInput, *args: Any, **kwargs: Any) -> str:
+        return str(input)
 
-    def _default_get_response(self, example: Example, experiment_run: ExperimentRun) -> str:
-        assert experiment_run.output is not None
-        return str(_unwrap_json(experiment_run.output.result))
+    def _default_get_response(
+        self, output: Optional[TaskOutput] = None, *args: Any, **kwargs: Any
+    ) -> str:
+        assert output is not None
+        return str(unwrap_json(output))
 
-    def evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        formatted_template = self._format_eval_template(example, exp_run)
+    def evaluate(
+        self,
+        *,
+        output: Optional[TaskOutput] = None,
+        metadata: ExampleMetadata = MappingProxyType({}),
+        input: ExampleInput = MappingProxyType({}),
+        **_: Any,
+    ) -> EvaluationResult:
+        formatted_template = self._format_eval_template(output, input, metadata)
         unparsed_response = self.model._generate(formatted_template)
         return self._parse_eval_output(unparsed_response)
 
-    async def async_evaluate(self, example: Example, exp_run: ExperimentRun) -> EvaluationResult:
-        formatted_template = self._format_eval_template(example, exp_run)
+    async def async_evaluate(
+        self,
+        *,
+        output: Optional[TaskOutput] = None,
+        metadata: ExampleMetadata = MappingProxyType({}),
+        input: ExampleInput = MappingProxyType({}),
+        **_: Any,
+    ) -> EvaluationResult:
+        formatted_template = self._format_eval_template(output, input, metadata)
         unparsed_response = await self.model._async_generate(formatted_template)
         return self._parse_eval_output(unparsed_response)

phoenix/experiments/evaluators/utils.py

@@ -0,0 +1,189 @@
+import functools
+import inspect
+from itertools import chain, islice, repeat
+from typing import TYPE_CHECKING, Any, Callable, Optional, Union
+
+from phoenix.experiments.types import (
+    AnnotatorKind,
+    EvaluationResult,
+    JSONSerializable,
+)
+
+if TYPE_CHECKING:
+    from phoenix.experiments.evaluators.base import Evaluator
+
+
+def unwrap_json(obj: JSONSerializable) -> JSONSerializable:
+    if isinstance(obj, dict):
+        if len(obj) == 1:
+            key = next(iter(obj.keys()))
+            output = obj[key]
+            assert isinstance(
+                output, (dict, list, str, int, float, bool, type(None))
+            ), "Output must be JSON serializable"
+            return output
+    return obj
+
+
+def validate_signature(sig: inspect.Signature) -> None:
+    # Check that the wrapped function has a valid signature for use as an evaluator
+    # If it does not, raise an error to exit early before running evaluations
+    params = sig.parameters
+    valid_named_params = {"input", "output", "expected", "metadata"}
+    if len(params) == 0:
+        raise ValueError("Evaluation function must have at least one parameter.")
+    if len(params) > 1:
+        for not_found in set(params) - valid_named_params:
+            param = params[not_found]
+            if (
+                param.kind is inspect.Parameter.VAR_KEYWORD
+                or param.default is not inspect.Parameter.empty
+            ):
+                continue
+            raise ValueError(
+                (
+                    f"Invalid parameter names in evaluation function: {', '.join(not_found)}. "
+                    "Parameters names for multi-argument functions must be "
+                    f"any of: {', '.join(valid_named_params)}."
+                )
+            )
+
+
+def _bind_signature(sig: inspect.Signature, **kwargs: Any) -> inspect.BoundArguments:
+    parameter_mapping = {
+        "input": kwargs.get("input"),
+        "output": kwargs.get("output"),
+        "expected": kwargs.get("expected"),
+        "metadata": kwargs.get("metadata"),
+    }
+    params = sig.parameters
+    if len(params) == 1:
+        parameter_name = next(iter(params))
+        if parameter_name in parameter_mapping:
+            return sig.bind(parameter_mapping[parameter_name])
+        else:
+            return sig.bind(parameter_mapping["output"])
+    return sig.bind_partial(
+        **{name: parameter_mapping[name] for name in set(parameter_mapping).intersection(params)}
+    )
+
+
+def create_evaluator(
+    kind: Union[str, AnnotatorKind] = AnnotatorKind.CODE,
+    name: Optional[str] = None,
+    scorer: Optional[Callable[[Any], EvaluationResult]] = None,
+) -> Callable[[Callable[..., Any]], "Evaluator"]:
+    if scorer is None:
+        scorer = _default_eval_scorer
+
+    if isinstance(kind, str):
+        kind = AnnotatorKind(kind.upper())
+
+    def wrapper(func: Callable[..., Any]) -> "Evaluator":
+        nonlocal name
+        if not name:
+            if hasattr(func, "__self__"):
+                name = func.__self__.__class__.__name__
+            elif hasattr(func, "__name__"):
+                name = func.__name__
+            else:
+                name = str(func)
+        assert name is not None
+
+        wrapped_signature = inspect.signature(func)
+        validate_signature(wrapped_signature)
+
+        if inspect.iscoroutinefunction(func):
+            return _wrap_coroutine_evaluation_function(name, kind, wrapped_signature, scorer)(func)
+        else:
+            return _wrap_sync_evaluation_function(name, kind, wrapped_signature, scorer)(func)
+
+    return wrapper
+
+
+def _wrap_coroutine_evaluation_function(
+    name: str,
+    annotator_kind: AnnotatorKind,
+    sig: inspect.Signature,
+    convert_to_score: Callable[[Any], EvaluationResult],
+) -> Callable[[Callable[..., Any]], "Evaluator"]:
+    from phoenix.experiments.evaluators.base import Evaluator
+
+    def wrapper(func: Callable[..., Any]) -> "Evaluator":
+        class AsyncEvaluator(Evaluator):
+            def __init__(self) -> None:
+                self._name = name
+                self._kind = annotator_kind
+
+            @functools.wraps(func)
+            async def __call__(self, *args: Any, **kwargs: Any) -> Any:
+                return await func(*args, **kwargs)
+
+            async def async_evaluate(self, **kwargs: Any) -> EvaluationResult:
+                bound_signature = _bind_signature(sig, **kwargs)
+                result = await func(*bound_signature.args, **bound_signature.kwargs)
+                return convert_to_score(result)
+
+        return AsyncEvaluator()
+
+    return wrapper
+
+
+def _wrap_sync_evaluation_function(
+    name: str,
+    annotator_kind: AnnotatorKind,
+    sig: inspect.Signature,
+    convert_to_score: Callable[[Any], EvaluationResult],
+) -> Callable[[Callable[..., Any]], "Evaluator"]:
+    from phoenix.experiments.evaluators.base import Evaluator
+
+    def wrapper(func: Callable[..., Any]) -> "Evaluator":
+        class SyncEvaluator(Evaluator):
+            def __init__(self) -> None:
+                self._name = name
+                self._kind = annotator_kind
+
+            @functools.wraps(func)
+            def __call__(self, *args: Any, **kwargs: Any) -> Any:
+                return func(*args, **kwargs)
+
+            def evaluate(self, **kwargs: Any) -> EvaluationResult:
+                bound_signature = _bind_signature(sig, **kwargs)
+                result = func(*bound_signature.args, **bound_signature.kwargs)
+                return convert_to_score(result)
+
+        return SyncEvaluator()
+
+    return wrapper
+
+
+def _default_eval_scorer(result: Any) -> EvaluationResult:
+    if isinstance(result, EvaluationResult):
+        return result
+    if isinstance(result, bool):
+        return EvaluationResult(score=float(result), label=str(result))
+    if hasattr(result, "__float__"):
+        return EvaluationResult(score=float(result))
+    if isinstance(result, str):
+        return EvaluationResult(label=result)
+    if isinstance(result, (tuple, list)) and 0 < len(result) <= 3:
+        # Possible interpretations are:
+        # - 3-tuple: (Score, Label, Explanation)
+        # - 2-tuple: (Score, Explanation) or (Label, Explanation)
+        # - 1-tuple: (Score, ) or (Label, )
+        # Note that (Score, Label) conflicts with (Score, Explanation) and we
+        # pick the latter because it's probably more prevalent. To get
+        # (Score, Label), use a 3-tuple instead, i.e. (Score, Label, None).
+        a, b, c = islice(chain(result, repeat(None)), 3)
+        score, label, explanation = None, a, b
+        if hasattr(a, "__float__"):
+            try:
+                score = float(a)
+            except ValueError:
+                pass
+            else:
+                label, explanation = (None, b) if len(result) < 3 else (b, c)
+        return EvaluationResult(score=score, label=label, explanation=explanation)
+    if result is None:
+        return EvaluationResult(score=0)
+    raise ValueError(f"Unsupported evaluation result type: {type(result)}")