Embedding Similarity

What It Is

Embedding similarity scores how semantically close two texts are by embedding each into a vector and taking the cosine similarity. In semantic-similarity-rating, a persona's free-text reaction is compared against anchor-statements for each scale level, and the resulting similarities become a probability distribution over the Likert scale.

What the Paper Found

  • Model: OpenAI text-embedding-3-small, cosine similarity. Upgrading to text-embedding-3-large changed nothing — the small model is sufficient.
  • Cost: embeddings are under a cent per test — negligible vs. the LLM generation tokens (~99% of marginal compute).
  • Caveat: the authors note results "depend on embedding-model choice" in general, even though the small/large swap didn't matter here — one input to domain-transfer-risk.

How It Applies to Marketing Factory

Embedding similarity is a cheap, general-purpose primitive the factory can reuse well beyond SSR — deduping content, clustering customer-voice verbatims, matching drafts to brand-voice exemplars, routing. For the SSR gate specifically, default to a small embedding model; the spend that matters is generation, not embedding.

Referenced from: ssr-paper-arxiv-2510-08338