Source: Qdrant indexing documentation

Qdrant's reference for payload indexes, HNSW parameters, filterable HNSW, the query planner, ACORN, sparse vector indexes and tenant-aware indexing.

Contents

Summary

Qdrant combines an HNSW vector index with typed payload indexes. Its distinctive move is filterable HNSW: once a field is indexed, extra graph edges are added per payload value so filtered traversals stay connected. A query planner uses payload-index cardinality estimates to pick between filtered graph search and a full scan over matching points.

Key claims

  • HNSW defaults: m 16, ef_construct 100; search ef defaults to ef_construct; configurable per collection or named vector.
  • full_scan_threshold (default 10,000 KB): if the estimated matching set is smaller, the planner does a full scan instead of graph traversal.
  • Filter-aware edges exist only if payload indexes are created before the HNSW index is built; create them right after the collection, before ingest. Otherwise a costly rebuild is needed.
  • ACORN (v1.16, released November 2025; enabled per query with the acorn search parameter): explores neighbors of neighbors for strict filter combinations or many deleted points, trading speed for accuracy.
  • Payload index types: keyword, integer, float, bool, geo, datetime, text (full text), uuid.
  • is_tenant: true on keyword/uuid fields co-locates a tenant’s data on disk.
  • Sparse vectors use an exact inverted-index-like structure with dot product only, optionally with an IDF modifier (v1.10) for BM25-style scoring.

Feeds

Vector databases · ANN indexes

Suggest changes →