langchain/docs/extras/guides/evaluation/string/embedding_distance.ipynb

Embedding Distance

To measure semantic similarity (or dissimilarity) between a prediction and a reference label string, you could use a vector vector distance metric the two embedded representations using the embedding_distance evaluator.^[1]

Note: This returns a distance score, meaning that the lower the number, the more similar the prediction is to the reference, according to their embedded representation.

Check out the reference docs for the EmbeddingDistanceEvalChain for more info.

In [1]:

from langchain.evaluation import load_evaluator

evaluator = load_evaluator("embedding_distance")

In [2]:

evaluator.evaluate_strings(prediction="I shall go", reference="I shan't go")

Out [2]:

{'score': 0.0966466944859925}

In [3]:

evaluator.evaluate_strings(prediction="I shall go", reference="I will go")

Out [3]:

{'score': 0.03761174337464557}

Select the Distance Metric

By default, the evalutor uses cosine distance. You can choose a different distance metric if you'd like.

In [4]:

from langchain.evaluation import EmbeddingDistance

list(EmbeddingDistance)

Out [4]:

[<EmbeddingDistance.COSINE: 'cosine'>,
 <EmbeddingDistance.EUCLIDEAN: 'euclidean'>,
 <EmbeddingDistance.MANHATTAN: 'manhattan'>,
 <EmbeddingDistance.CHEBYSHEV: 'chebyshev'>,
 <EmbeddingDistance.HAMMING: 'hamming'>]

In [5]:

# You can load by enum or by raw python string
evaluator = load_evaluator(
    "embedding_distance", distance_metric=EmbeddingDistance.EUCLIDEAN
)

Select Embeddings to Use

The constructor uses OpenAI embeddings by default, but you can configure this however you want. Below, use huggingface local embeddings

In [6]:

from langchain.embeddings import HuggingFaceEmbeddings

embedding_model = HuggingFaceEmbeddings()
hf_evaluator = load_evaluator("embedding_distance", embeddings=embedding_model)

In [7]:

hf_evaluator.evaluate_strings(prediction="I shall go", reference="I shan't go")

Out [7]:

{'score': 0.5486443280477362}

In [8]:

hf_evaluator.evaluate_strings(prediction="I shall go", reference="I will go")

Out [8]:

{'score': 0.21018880025138598}

1. Note: When it comes to semantic similarity, this often gives better results than older string distance metrics (such as those in the StringDistanceEvalChain), though it tends to be less reliable than evaluators that use the LLM directly (such as the QAEvalChain or LabeledCriteriaEvalChain)