Hybrid Search

Hybrid search combines dense (semantic) and sparse (BM25 keyword) retrieval using Reciprocal Rank Fusion (RRF) or Distribution-Based Score Fusion (DBSF).

Basic Hybrid Search

QUERY 'emergency care' FROM docs LIMIT 5 USING HYBRID

This runs dense and sparse queries in parallel and merges results with RRF.

Fusion Modes

RRF (default):

QUERY 'vector database' FROM docs LIMIT 10 USING HYBRID

DBSF:

QUERY 'vector database' FROM docs LIMIT 10 USING HYBRID FUSION DBSF

Parameterized RRF:

QUERY 'vector database' FROM docs LIMIT 10 USING HYBRID WITH (rrf_k = 30, rrf_weights = [0.7, 0.3])

Parameter	Default	Description
rrf_k	60	Rank penalty constant — lower values make top-ranked items dominate more
rrf_weights	[0.5, 0.5]	Per-leg weights — [dense_weight, sparse_weight]

With Filters

QUERY 'emergency care' FROM docs LIMIT 10
  USING HYBRID
  WHERE status = 'active' AND year >= 2024

Always Create Payload Indexes

Always create payload indexes before filtering. Without an index, Qdrant performs a full scan:

CREATE INDEX ON docs FOR status TYPE keyword
CREATE INDEX ON docs FOR year TYPE integer

MMR Diversity

Apply Maximal Marginal Relevance to reduce redundant results in the dense leg:

QUERY 'emergency care' FROM docs LIMIT 10
  USING HYBRID
  WITH (mmr_diversity = 0.5, mmr_candidates = 100)

Parameter	Description
mmr_diversity	0..1 — higher = more diverse
mmr_candidates	Candidate pool size before MMR selection

Dense-Only / Sparse-Only

Dense only:

QUERY 'vector database' FROM docs LIMIT 10 USING DENSE

Sparse (BM25) only:

QUERY 'vector database' FROM docs LIMIT 10 USING SPARSE

Reranking (Cloud)

Apply cross-encoder reranking after hybrid retrieval (Qdrant Cloud only):

QUERY 'emergency care' FROM docs LIMIT 10 USING HYBRID RERANK

Multi-Stage Hybrid with CTEs

For fine-grained control over each retrieval leg:

WITH
  dense AS (QUERY 'vector database performance' USING dense LIMIT 200
    WHERE category = 'tech' AND published_at >= '2025-01-01'),
  sparse AS (QUERY 'vector database performance' USING sparse LIMIT 300)
QUERY 'vector database performance' FROM articles LIMIT 10
  PREFETCH (
    dense SCORE THRESHOLD 0.6,
    sparse SCORE THRESHOLD 0.3
  )
  FUSION RRF
  WITH (rrf_k = 20, rrf_weights = [0.6, 0.4])

Key decisions

• Dense leg has a filter for recent tech articles — higher precision.
• Sparse leg is wider — catches exact terminology matches.
• rrf_weights = [0.6, 0.4] favors the semantic leg.
• rrf_k = 20 (vs default 60) makes rank penalties steeper.

For full CTE patterns, see the Multi-Stage Retrieval guide.

Tuned RRF

Production hybrid search with per-prefetch filters and weighted RRF fusion:

WITH
  dense AS (
    QUERY 'vector database performance' USING dense LIMIT 200
    WHERE category = 'tech' AND published_at >= '2025-01-01'
  ),
  sparse AS (
    QUERY 'vector database performance' USING sparse LIMIT 300
  )
QUERY 'vector database performance' FROM articles LIMIT 10
  PREFETCH (
    dense SCORE THRESHOLD 0.6,
    sparse SCORE THRESHOLD 0.3
  )
  FUSION RRF
  WITH (rrf_k = 20, rrf_weights = [0.6, 0.4])

Decision	Why
Dense prefetch filtered to recent tech	Higher precision leg
Sparse prefetch has no filter	Wider keyword net
rrf_weights = [0.6, 0.4]	Favors the semantic leg
rrf_k = 20 (default: 60)	Steeper rank penalty — top-ranked items dominate more
SCORE THRESHOLD per leg	Prune low-quality candidates before fusion

Complete Clause Reference

QUERY '<text>' FROM <collection>
  [USING HYBRID [FUSION DBSF]]
  [WITH (rrf_k = <n>, rrf_weights = [<f>, <f>])]
  [WITH (mmr_diversity = <f>, mmr_candidates = <n>)]
  [WITH (hnsw_ef = <n>)]
  [WITH (acorn = true)]
  [WITH (exact = true)]
  [WHERE <filter>]
  [RERANK]
  [GROUP BY '<field>' [GROUP_SIZE <n>]]
  [SCORE THRESHOLD <float>]
  [LIMIT <n>] [OFFSET <n>]