Named vectors let one Qdrant point contain multiple independently configured vector spaces. For example, a product point can have a text embedding and an image embedding, each with its own dimensionality and distance metric.

This is useful when the vectors represent different modalities, fields, or retrieval models but should still belong to the same logical entity. Qdrant's current model also distinguishes named vectors from multivectors: a named vector is a separate vector field, while a multivector represents a collection or matrix of vectors used by models such as late-interaction retrieval.

The trade-off is simplicity versus flexibility. Separate collections can isolate workloads and schemas more strongly, but they require application-level joins between results. Named vectors keep the logical entity together, at the cost of storing and maintaining more vector data per point.

A common misconception is that 'multiple vectors' automatically means late interaction. They are different concepts. Named vectors are separate vector spaces; multivectors are a specific vector representation supported by Qdrant.

javascript

1
Named vectors keep multiple independently configured vector spaces on one point
2
Different named vectors can have different dimensions and metrics
3
Named vectors are distinct from multivectors used by late-interaction approaches
4
Separate collections can be preferable when workloads require strong operational isolation

Difficulty: 5/10

Topics: Named vectors, Multivectors

Scenario Questions

0-2 years experience

1
A product has both a text description and an image. Why might storing two named vectors be better than concatenating the vectors?
2
Your application only searches product text. Would adding an image vector to every point automatically improve text search?

2-5 years experience

1
You need to replace only the image embedding model while keeping text search unchanged. How would named vectors help with the migration?
2
A team wants to combine title and body embeddings as two named vectors but expects one nearest-neighbor ranking across both. What design issue would you raise?

5-8 years experience

1
You are building multimodal product retrieval where users may query by text or image. How would you design the Qdrant schema and query paths?
2
Storage costs have doubled because each point now carries four vector spaces. How would you decide which representations should remain in Qdrant?

8+ years experience

1
A platform supports dozens of embedding models and modalities. How would you prevent named-vector schema growth from becoming an operational liability?
2
You need late-interaction retrieval alongside conventional dense retrieval. How would you distinguish the data model and evaluation strategy for named vectors versus multivectors?

Follow-up Questions

❓When would you choose separate collections instead of named vectors?
❓How do named vectors differ from multivectors?