Module haystack_experimental.evaluation.harness.evaluation_harness

EvaluationRunOverrides

Overrides for an evaluation run.

Used to override the init parameters of components in either
(or both) the evaluated and evaluation pipelines. Each key is
a component name and its value a dictionary with init parameters
to override.

Arguments:

• evaluated_pipeline_overrides: Overrides for the evaluated pipeline.
• evaluation_pipeline_overrides: Overrides for the evaluation pipeline.

EvaluationHarness

Executes a pipeline with a given set of parameters, inputs and evaluates its outputs with an evaluation pipeline.

EvaluationHarness.run

@abstractmethod
def run(inputs: EvalRunInputT,
        *,
        overrides: Optional[EvalRunOverridesT] = None,
        run_name: Optional[str] = None) -> EvalRunOutputT

Launch a evaluation run.

Arguments:

• inputs: Inputs to the evaluated and evaluation pipelines.
• overrides: Overrides for the harness.
• run_name: A name for the evaluation run.

Returns:

The output of the evaluation pipeline.

Module haystack_experimental.evaluation.harness.rag.harness

DefaultRAGArchitecture

Represents default RAG pipeline architectures that can be used with the evaluation harness.

EMBEDDING_RETRIEVAL

A RAG pipeline with:

• A query embedder component named 'query_embedder' with a 'text' input.
• A document retriever component named 'retriever' with a 'documents' output.

KEYWORD_RETRIEVAL

A RAG pipeline with:

• A document retriever component named 'retriever' with a 'query' input and a 'documents' output.

GENERATION_WITH_EMBEDDING_RETRIEVAL

A RAG pipeline with:

• A query embedder component named 'query_embedder' with a 'text' input.
• A document retriever component named 'retriever' with a 'documents' output.
• A response generator component named 'generator' with a 'replies' output.

GENERATION_WITH_KEYWORD_RETRIEVAL

A RAG pipeline with:

• A document retriever component named 'retriever' with a 'query' input and a 'documents' output.
• A response generator component named 'generator' with a 'replies' output.

DefaultRAGArchitecture.expected_components

@property
def expected_components(
) -> Dict[RAGExpectedComponent, RAGExpectedComponentMetadata]

Returns the expected components for the architecture.

Returns:

The expected components.

RAGEvaluationHarness

Evaluation harness for evaluating RAG pipelines.

RAGEvaluationHarness.__init__

def __init__(rag_pipeline: Pipeline,
             rag_components: Union[
                 DefaultRAGArchitecture,
                 Dict[RAGExpectedComponent, RAGExpectedComponentMetadata],
             ],
             metrics: Set[RAGEvaluationMetric],
             *,
             progress_bar: bool = True)

Create an evaluation harness for evaluating basic RAG pipelines.

Arguments:

• rag_pipeline: The RAG pipeline to evaluate.
• rag_components: Either a default RAG architecture or a mapping
  of expected components to their metadata.
• metrics: The metrics to use during evaluation.
• progress_bar: Whether to display a progress bar during evaluation.

Module haystack_experimental.evaluation.harness.rag.parameters

RAGExpectedComponent

Represents the basic components in a RAG pipeline that are, by default, required to be present for evaluation.

Each of these can be separate components in the pipeline or a single component that performs
multiple tasks.

QUERY_PROCESSOR

The component in a RAG pipeline that accepts the user query.
Expected inputs: query - Name of input that contains the query string.

DOCUMENT_RETRIEVER

The component in a RAG pipeline that retrieves documents based on the query.
Expected outputs: retrieved_documents - Name of output containing retrieved documents.

RESPONSE_GENERATOR

The component in a RAG pipeline that generates responses based on the query and the retrieved documents.
Can be optional if the harness is only evaluating retrieval.
Expected outputs: replies - Name of out containing the LLM responses. Only the first response is used.

RAGExpectedComponentMetadata

Metadata for a RAGExpectedComponent.

Arguments:

• name: Name of the component in the pipeline.
• input_mapping: Mapping of the expected inputs to
  corresponding component input names.
• output_mapping: Mapping of the expected outputs to
  corresponding component output names.

RAGEvaluationMetric

Represents the metrics that can be used to evaluate a RAG pipeline.

DOCUMENT_MAP

Document Mean Average Precision.
Required RAG components: Query Processor, Document Retriever.

DOCUMENT_MRR

Document Mean Reciprocal Rank.
Required RAG components: Query Processor, Document Retriever.

DOCUMENT_RECALL_SINGLE_HIT

Document Recall with a single hit.
Required RAG components: Query Processor, Document Retriever.

DOCUMENT_RECALL_MULTI_HIT

Document Recall with multiple hits.
Required RAG components: Query Processor, Document Retriever.

SEMANTIC_ANSWER_SIMILARITY

Semantic Answer Similarity.
Required RAG components: Query Processor, Response Generator.

FAITHFULNESS

Faithfulness.
Required RAG components: Query Processor, Document Retriever, Response Generator.

CONTEXT_RELEVANCE

Context Relevance.
Required RAG components: Query Processor, Document Retriever.

RAGEvaluationInput

Input passed to the RAG evaluation harness.

Arguments:

• queries: The queries passed to the RAG pipeline.
• ground_truth_documents: The ground truth documents passed to the
  evaluation pipeline. Only required for metrics
  that require them. Corresponds to the queries.
• ground_truth_answers: The ground truth answers passed to the
  evaluation pipeline. Only required for metrics
  that require them. Corresponds to the queries.
• rag_pipeline_inputs: Additional inputs to pass to the RAG pipeline. Each
  key is the name of the component and its value a dictionary
  with the input name and a list of values, each corresponding
  to a query.

RAGEvaluationOverrides

Overrides for a RAG evaluation run.

Used to override the init parameters of components in
either (or both) the evaluated and evaluation pipelines.

Arguments:

• rag_pipeline: Overrides for the RAG pipeline. Each
  key is a component name and its value a dictionary
  with init parameters to override.
• eval_pipeline: Overrides for the evaluation pipeline. Each
  key is a RAG metric and its value a dictionary
  with init parameters to override.

RAGEvaluationOutput

Represents the output of a RAG evaluation run.

Arguments:

• evaluated_pipeline: Serialized version of the evaluated pipeline, including overrides.
• evaluation_pipeline: Serialized version of the evaluation pipeline, including overrides.
• inputs: Input passed to the evaluation harness.
• results: Results of the evaluation run.