Skip to main content

Asymmetric Benchmark

Validate that Voyage 4 models produce compatible embeddings by measuring cross-model similarity.

Run​

vai benchmark asymmetric

What It Measures​

  • Embeds the same texts with different Voyage 4 models
  • Computes cosine similarity between embeddings from different models
  • Shows how well the shared embedding space works in practice

Expected Results​

Embeddings from different Voyage 4 models should have high cosine similarity (typically 0.95+) for the same input text. This confirms that asymmetric retrieval — embedding documents with one model and queries with another — works reliably.

Why It Matters​

If cross-model similarity is high, you can safely:

  • Embed documents with voyage-4-lite ($0.02/1M) for cost savings
  • Embed queries with voyage-4-large ($0.12/1M) for quality
  • Mix and match without re-embedding your corpus

Further Reading​