RunEvalConfig

class langchain.smith.evaluation.config.RunEvalConfig

Bases: BaseModel

Configuration for a run evaluation.

Parameters:

• evaluators (List[Union[EvaluatorType, EvalConfig, RunEvaluator, Callable]]) – Configurations for which evaluators to apply to the dataset run. Each can be the string of an EvaluatorType, such as EvaluatorType.QA, the evaluator type string (“qa”), or a configuration for a given evaluator (e.g., RunEvalConfig.QA).

• custom_evaluators (Optional[List[Union[RunEvaluator, StringEvaluator]]]) – Custom evaluators to apply to the dataset run.

• reference_key (Optional[str]) – The key in the dataset run to use as the reference string. If not provided, it will be inferred automatically.

• prediction_key (Optional[str]) – The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

• input_key (Optional[str]) – The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

• eval_llm (Optional[BaseLanguageModel]) – The language model to pass to any evaluators that use a language model.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param batch_evaluators: List[Callable[[Sequence[Run], Sequence[Example] | None], EvaluationResult | EvaluationResults | dict]] | None = None

Evaluators that run on an aggregate/batch level.

These generate 1 or more metrics that are assigned to the full test run. As a result, they are not associated with individual traces.

param custom_evaluators: List[Callable[[Run, Example | None], EvaluationResult | EvaluationResults | dict] | RunEvaluator | StringEvaluator] | None = None

Custom evaluators to apply to the dataset run.

param eval_llm: BaseLanguageModel | None = None

The language model to pass to any evaluators that require one.

param evaluators: List[EvaluatorType | str | EvalConfig | Callable[[Run, Example | None], EvaluationResult | EvaluationResults | dict] | RunEvaluator | StringEvaluator] [Optional]

Configurations for which evaluators to apply to the dataset run. Each can be the string of an EvaluatorType, such as EvaluatorType.QA, the evaluator type string (“qa”), or a configuration for a given evaluator (e.g., RunEvalConfig.QA).

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

class CoTQA

Bases: SingleKeyEvalConfig

Configuration for a context-based QA evaluator.

Parameters:

• prompt (Optional[BasePromptTemplate]) – The prompt template to use for generating the question.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.CONTEXT_QA

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param prompt: BasePromptTemplate | None = None

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class ContextQA

Bases: SingleKeyEvalConfig

Configuration for a context-based QA evaluator.

Parameters:

• prompt (Optional[BasePromptTemplate]) – The prompt template to use for generating the question.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.CONTEXT_QA

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param prompt: BasePromptTemplate | None = None

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class Criteria

Bases: SingleKeyEvalConfig

Configuration for a reference-free criteria evaluator.

Parameters:

• criteria (Optional[CRITERIA_TYPE]) – The criteria to evaluate.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param criteria: Mapping[str, str] | Criteria | ConstitutionalPrinciple | None = None

param evaluator_type: EvaluatorType = EvaluatorType.CRITERIA

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class EmbeddingDistance

Bases: SingleKeyEvalConfig

Configuration for an embedding distance evaluator.

Parameters:

• embeddings (Optional[Embeddings]) – The embeddings to use for computing the distance.

• distance_metric (Optional[EmbeddingDistanceEnum]) – The distance metric to use for computing the distance.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param distance_metric: EmbeddingDistance | None = None

param embeddings: Embeddings | None = None

param evaluator_type: EvaluatorType = EvaluatorType.EMBEDDING_DISTANCE

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class ExactMatch

Bases: SingleKeyEvalConfig

Configuration for an exact match string evaluator.

Parameters:

• ignore_case (bool) – Whether to ignore case when comparing strings.

• ignore_punctuation (bool) – Whether to ignore punctuation when comparing strings.

• ignore_numbers (bool) – Whether to ignore numbers when comparing strings.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.EXACT_MATCH

param ignore_case: bool = False

param ignore_numbers: bool = False

param ignore_punctuation: bool = False

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class JsonEqualityEvaluator

Bases: EvalConfig

Configuration for a json equality evaluator.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.JSON_EQUALITY

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class JsonValidity

Bases: SingleKeyEvalConfig

Configuration for a json validity evaluator.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.JSON_VALIDITY

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class LabeledCriteria

Bases: SingleKeyEvalConfig

Configuration for a labeled (with references) criteria evaluator.

Parameters:

• criteria (Optional[CRITERIA_TYPE]) – The criteria to evaluate.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param criteria: Mapping[str, str] | Criteria | ConstitutionalPrinciple | None = None

param evaluator_type: EvaluatorType = EvaluatorType.LABELED_CRITERIA

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class LabeledScoreString

Bases: ScoreString

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param criteria: CRITERIA_TYPE | None = None

param evaluator_type: EvaluatorType = EvaluatorType.LABELED_SCORE_STRING

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param normalize_by: float | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param prompt: BasePromptTemplate | None = None

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class QA

Bases: SingleKeyEvalConfig

Configuration for a QA evaluator.

Parameters:

• prompt (Optional[BasePromptTemplate]) – The prompt template to use for generating the question.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.QA

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param prompt: BasePromptTemplate | None = None

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class RegexMatch

Bases: SingleKeyEvalConfig

Configuration for a regex match string evaluator.

Parameters:

flags (int) – The flags to pass to the regex. Example: re.IGNORECASE.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param evaluator_type: EvaluatorType = EvaluatorType.REGEX_MATCH

param flags: int = 0

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class ScoreString

Bases: SingleKeyEvalConfig

Configuration for a score string evaluator. This is like the criteria evaluator but it is configured by default to return a score on the scale from 1-10.

It is recommended to normalize these scores by setting normalize_by to 10.

Parameters:

• criteria (Optional[CRITERIA_TYPE]) – The criteria to evaluate.

• llm (Optional[BaseLanguageModel]) – The language model to use for the evaluation chain.

• normalize_by (Optional[int] = None) – If you want to normalize the score, the denominator to use. If not provided, the score will be between 1 and 10 (by default).

• prompt (Optional[BasePromptTemplate])

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param criteria: Mapping[str, str] | Criteria | ConstitutionalPrinciple | None = None

param evaluator_type: EvaluatorType = EvaluatorType.SCORE_STRING

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param llm: BaseLanguageModel | None = None

param normalize_by: float | None = None

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param prompt: BasePromptTemplate | None = None

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]

class StringDistance

Bases: SingleKeyEvalConfig

Configuration for a string distance evaluator.

Parameters:

distance (Optional[StringDistanceEnum]) – The string distance metric to use.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

param distance: StringDistance | None = None

The string distance metric to use. damerau_levenshtein: The Damerau-Levenshtein distance. levenshtein: The Levenshtein distance. jaro: The Jaro distance. jaro_winkler: The Jaro-Winkler distance.

param evaluator_type: EvaluatorType = EvaluatorType.STRING_DISTANCE

param input_key: str | None = None

The key from the traced run’s inputs dictionary to use to represent the input. If not provided, it will be inferred automatically.

param normalize_score: bool = True

Whether to normalize the distance to between 0 and 1. Applies only to the Levenshtein and Damerau-Levenshtein distances.

param prediction_key: str | None = None

The key from the traced run’s outputs dictionary to use to represent the prediction. If not provided, it will be inferred automatically.

param reference_key: str | None = None

The key in the dataset run to use as the reference string. If not provided, we will attempt to infer automatically.

get_kwargs() → Dict[str, Any]

Get the keyword arguments for the load_evaluator call.

Returns:

The keyword arguments for the load_evaluator call.

Return type:

Dict[str, Any]