Embeddings

EmbeddingModel calls record and replay too, independently of chat — enable with spring.ai.test.vcr.embedding.enabled=true (separate from, and off unless enabled alongside, the top-level spring.ai.test.vcr.enabled):

spring:
  ai:
    test:
      vcr:
        embedding:
          enabled: true
          mode: RECORD_OR_REPLAY
          cache-directory: src/test/resources/llm-cache-embedding

EmbeddingModel has no advisor chain the way ChatClient does, so interception wraps the EmbeddingModel bean itself, transparently — @Autowired EmbeddingModel in your own code is unchanged, and every entry point (embed(String), embed(List<String>), embedForResponse(List<String>)) is covered, since they all route through the same underlying call.

@Autowired
private EmbeddingModel embeddingModel;

@Test
void replaysTheExactRecordedVector() {
    float[] first = this.embeddingModel.embed("Istanbul is the largest city in Turkey.");
    float[] second = this.embeddingModel.embed("Istanbul is the largest city in Turkey.");

    assertThat(second).isEqualTo(first); // exactly, not "same length"
}

A replayed vector is exactly — not "the same length as" — what was recorded.

Why this exists on its own

This is the foundation Assertions' semantic similarity check depends on: an assertion that computes cosine similarity against a reference answer needs its own embedding call to be exactly this deterministic, or every CI run would make a live, non-reproducible embedding call to check a "deterministic" test.