Database / Qdrant Vector DB Interview questions
What distance metrics does Qdrant support?
Qdrant supports several standard distance/similarity metrics for comparing vectors, chosen when a collection is created, and the right choice generally depends on how the embedding model that produced the vectors was trained and normalized.
| Metric | Typical Use |
| Cosine | Measures angle between vectors; common for text embeddings, ignores magnitude |
| Dot Product | Raw dot product; fast, often used when vectors are pre-normalized |
| Euclidean (L2) | Straight-line distance; common for spatial or image-based embeddings |
| Manhattan (L1) | Sum of absolute differences; less common, used in specific domains |
The metric is fixed per collection (or per named vector, if a collection uses several), and it must match what the embedding model producing the vectors was actually optimized for — using cosine distance with a model trained assuming dot-product similarity, for instance, can silently produce worse-quality search results even though the query still runs without any error.
Which metric is common for text embeddings and ignores vector magnitude?
What can happen if the chosen distance metric doesn't match what the embedding model was optimized for?
More Related questions...
