Implementation:Explodinggradients Ragas SemanticSimilarity Metric

Field	Value
source	Repo
domains	Metrics, Embeddings
last_updated	2026-02-10

Overview

SemanticSimilarity (also aliased as AnswerSimilarity) computes cosine similarity between the embeddings of a generated response and a reference ground truth.

Description

The SemanticSimilarity class scores the semantic similarity between a reference answer and a generated response using embedding-based cosine similarity. It supports both standard embedding models and cross-encoder models (via HuggingFace). The metric can optionally apply a threshold to produce binary output. It inherits from MetricWithEmbeddings and SingleTurnMetric.

Key attributes:

• is_cross_encoder -- Boolean indicating if a cross-encoder model is used (auto-detected from HuggingFace embeddings).
• threshold -- Optional float threshold for binary classification (default None).

The AnswerSimilarity subclass provides a backward-compatible alias with name="answer_similarity".

Usage

The metric requires reference and response columns. Embeddings must be configured.

Code Reference

Property	Value
Source Location	src/ragas/metrics/_answer_similarity.py L25-116
Class Signature	class SemanticSimilarity(MetricWithEmbeddings, SingleTurnMetric)
Import	from ragas.metrics import SemanticSimilarity

I/O Contract

Inputs

Parameter	Type	Required	Description
reference	str	Yes	The ground truth reference text
response	str	Yes	The generated response to evaluate

Outputs

Output	Type	Description
score	float	Cosine similarity score (0.0 to 1.0), or binary (0/1) if threshold is set

Usage Examples

from ragas.metrics import SemanticSimilarity
from ragas.dataset_schema import SingleTurnSample

metric = SemanticSimilarity(threshold=None)
# metric.embeddings = ...  # Set your embeddings

sample = SingleTurnSample(
    reference="The Eiffel Tower is located in Paris, France.",
    response="The Eiffel Tower can be found in Paris."
)
# score = await metric.single_turn_ascore(sample)

A pre-configured instance is available:

from ragas.metrics._answer_similarity import answer_similarity

Related Pages

• Implementation:Explodinggradients_Ragas_AnswerCorrectness_Metric -- Uses SemanticSimilarity as a component for scoring
• Implementation:Explodinggradients_Ragas_ResponseRelevancy_Metric -- Another embedding-based relevancy metric
• Implementation:Explodinggradients_Ragas_StringMetrics_Module -- Non-LLM string similarity alternatives