How should you choose a distance metric when creating a Qdrant collection?

I start with the embedding model's documentation or training objective. The metric is part of the representation's retrieval semantics, so I would not choose it solely because it is popular.

For normalized semantic embeddings, cosine is a common choice. If the model is explicitly optimized for dot product or magnitude carries useful information, dot product may be more appropriate. Euclidean or Manhattan can make sense for representations whose geometry was designed around those distances.

The trade-off is between following the model's intended geometry and optimizing the application's observed retrieval quality. If documentation is unclear, I would build an evaluation set with representative queries and compare metrics against human or labeled relevance judgments.

One important operational constraint is that the vector configuration is part of the collection schema. Changing the metric is not equivalent to changing a query parameter; if the representation semantics change, plan a migration or a separate vector space.