.svg)
What is Embedding Similarity Metric?
An Embedding Similarity Metric embeds texts into vectors and scores how close they are--often cosine similarity--so you can judge semantic relatedness without requiring exact string match.
What this looks like in production
Teams compare a model answer to a gold reference, or a query to retrieved chunks, to catch meaning drift that keyword metrics miss.
What teams usually do about it
- Pick an embedding model suited to your domain language.
- Validate thresholds against human judgments.
- Pair with groundedness checks when citations matter.
Related Giskard articles
Add semantic checks in Giskard
Pair embedding similarity with groundedness and retrieval checks in Giskard so semantic closeness is not mistaken for factual correctness. See OWASP LLM risks for related failure modes.
Further reading
Authority reference: Sentence-Transformers documentation.
