Cosine similarity and inner product are the similarity metrics that work best with text-embedding-3-small embeddings. These metrics are well suited for dense semantic vectors and are widely supported by vector databases. They measure how closely two vectors align in direction, which correlates well with semantic similarity.
Cosine similarity is the most common choice because it normalizes vector length and focuses purely on direction. This is useful when embedding magnitudes are not meaningful by themselves, which is usually the case for text embeddings. Inner product is also widely used and often performs similarly, especially when vectors are already normalized. Both metrics tend to produce stable, intuitive rankings for semantic search and recommendation use cases.
Vector databases such as Milvus and Zilliz Cloud support these metrics natively and allow developers to configure them per collection. Choosing the right metric is usually less important than ensuring consistency: the same metric should be used during indexing and querying. In practice, most teams start with cosine similarity and only experiment further if they see specific ranking issues. text-embedding-3-small is designed to work smoothly with these standard metrics, making integration straightforward.
For more information, click here： https://zilliz.com/ai-models/text-embedding-3-small