Can two points in the same Qdrant collection have vectors of different dimensionality?
Vector dimensions are defined by the vector space
For a given dense vector space in a Qdrant collection, points must conform to the configured dimensionality. A query vector must also match that vector space's dimension.
This constraint exists because the index and distance computation operate in a defined vector space. Allowing arbitrary dimensions within one vector field would make the metric and index structure ambiguous.
Named vectors provide the escape hatch when one logical point needs different dimensions: each named vector is a separate configured vector space. For example, an image vector can be 512 dimensions while a text vector is 768 dimensions on the same point.
A common mistake is assuming one collection means one dimension globally. That is true for a single unnamed dense vector configuration, but named vectors allow multiple independently configured vector spaces in the same collection.
- 1
A vector field has a fixed configured dimensionality
- 2
Query vectors must match the selected vector space
- 3
Named vectors can hold different dimensions on the same point
- 4
Changing an embedding dimension normally requires a migration or a new vector space
Scenario Questions
0-2 years experience
- 1Your model produces 384-dimensional vectors but the collection expects 768. What would you expect when inserting the vector?
- 2Why cannot a nearest-neighbor search directly compare a 384-dimensional vector with a 768-dimensional vector?
2-5 years experience
- 1You must support two embedding models with dimensions 384 and 1536 during a migration. How could you model this in Qdrant?
- 2A collection contains text vectors of 768 dimensions and a new model produces 1024 dimensions. Would you overwrite the existing vector field? Why or why not?
5-8 years experience
- 1You need a zero-downtime migration between embedding models with different dimensions across 500 million points. How would you stage the schema and backfill?
- 2A deployment accidentally sends vectors from two model versions into the same vector field. How would you detect and recover from the corruption?
8+ years experience
- 1Your organization frequently upgrades embedding models. What schema and deployment strategy would you establish to make dimensionality changes routine rather than disruptive?
- 2You need to retain multiple model versions for A/B testing while controlling vector-storage cost. How would you design the vector layout?
Follow-up Questions
- What happens if an inserted vector has the wrong dimension?
- How would you migrate a collection to a new embedding model with a different dimension?