What is Qdrant's Scroll API used for, and why is it preferable to vector search when iterating through matching points?

Scroll is designed for iterating through points, optionally constrained by a payload filter, rather than ranking the dataset by vector similarity. It is useful for exports, migrations, reindexing, cleanup jobs, and administrative scans.

Vector search answers a different question: which points are most similar to this query vector? If the goal is to visit every point matching a condition, repeatedly running similarity searches is both semantically wrong and computationally wasteful.

Scroll uses continuation state so the caller can request the next portion of the dataset. This makes it suitable for processing large collections without loading the entire result set into application memory.

A common mistake is using offset-like search pagination to walk a whole collection. Search is optimized for ranked retrieval, while scroll is designed for systematic traversal. Exact pagination and ordering details are version-dependent.