A point lookup addresses a known ID directly, so it is conceptually similar to a primary-key read. It does not calculate similarity or rank neighboring vectors.

Vector search starts with a query vector and asks Qdrant to rank candidate points according to the configured distance metric. The caller may not know the IDs beforehand. That distinction matters because direct lookup is appropriate when an application already knows the entity, while vector search is appropriate for discovery by semantic similarity.

The trade-off is deterministic identity lookup versus relevance-based retrieval. A common architecture uses vector search to discover candidate IDs and then direct database or point lookups to fetch authoritative application state.

A common mistake is using vector search to retrieve a known document. That wastes search resources and introduces unnecessary relevance logic.

javascript

1
ID lookup targets known point identities
2
Vector search ranks candidates by vector similarity
3
Direct lookup is deterministic and avoids ANN traversal
4
Search results can be followed by authoritative source-of-truth lookups

Difficulty: 2/10

Topics: Point retrieval, Vector search, CRUD operations

Scenario Questions

0-2 years experience

1
Your API receives a known Qdrant point ID from a previous search. Would you perform another vector search to fetch that point?
2
A user provides a document ID but the application performs a semantic search to retrieve it. What inefficiency would you identify?

2-5 years experience

1
Your search service returns Qdrant IDs but needs authoritative pricing from PostgreSQL. How would you combine the two lookup patterns?
2
A developer reports that direct point retrieval returns a document instantly while semantic search is slower. Why is that expected?

5-8 years experience

1
A vector search returns 100 candidate IDs and the application performs 100 sequential point lookups. What architecture or batching improvement would you consider?
2
Your authorization data is authoritative outside Qdrant. How would you design the retrieval flow so semantic search does not become an authorization bypass?

8+ years experience

1
You are designing a retrieval service that separates candidate generation from document hydration. Where would Qdrant fit, and why would you avoid using it as the sole source of truth?
2
Search traffic is high but most requests eventually need full application records from another database. How would you optimize the boundary between vector retrieval and entity hydration?

Follow-up Questions

❓When would you use retrieve instead of query_points?
❓Why might an application perform a vector search followed by a direct lookup?