Embeddings and reranking

Vector names answer "which vector?" Vector roles answer "what input shape and embedding path?" QQL keeps those decisions separate.

Form	Role
`USING semantic`	Resolve role from collection schema
`USING semantic AS DENSE`	Explicit dense vector
`USING lexical AS SPARSE`	Explicit sparse vector
`USING colbert AS MULTI`	Explicit dense multivector

QQLOffline-explicit vector roleTry in playground

QUERY 'late interaction' FROM papers USING colbert AS MULTI LIMIT 10;

Schema with MULTIVECTOR

Token-level multivector storage uses MULTIVECTOR with a comparator. The example below keeps a dense vector for candidate generation and a 128 dimension multivector for late interaction.

QQLCreate multi-vector collection for ColBERTTry in playground

CREATE COLLECTION kb_articles (
  dense VECTOR(384, COSINE),
  colbert VECTOR(128, COSINE) WITH HNSW (m = 0) WITH MULTIVECTOR (comparator = 'max_sim')
);

Direct multivector search

Search a multivector space directly with USING ... AS MULTI.

QQLDirect ColBERT token-level searchTry in playground

QUERY 'late interaction vector retrieval'
FROM kb_articles
USING colbert AS MULTI
WHERE category = 'search'
LIMIT 10;

Two-stage dense retrieval plus rerank

At scale, retrieve candidates with dense search, then rerank with late-interaction MaxSim over the prefetched set.

QQLTwo-stage prefetch and ColBERT rerankingTry in playground

WITH
  candidates AS (QUERY 'late interaction vector retrieval' USING dense LIMIT 100)
QUERY RERANK TEXT 'late interaction vector retrieval' MODEL 'colbert-v2'
FROM kb_articles
USING colbert AS MULTI
PREFETCH (candidates)
LIMIT 10;

Cross-encoder reranking

QQLHost pair scoringTry in playground

WITH
  candidates AS (QUERY 'vector compression' USING dense LIMIT 50)
QUERY CROSS RERANK TEXT 'vector compression' MODEL 'bge-reranker-base' ON FIELD abstract
FROM papers
PREFETCH (candidates)
LIMIT 10;

Cross reranking reads candidate text from the payload, calls the host pair scoring capability, and reorders hits client-side. It does not use a Qdrant multivector target.

Architecture

Architecture	Retrieval latency	Storage footprint	Interaction depth	Best used for
Single dense vector	1-5 ms	Low (1 vector per doc)	Coarse chunk level	First-stage candidate generation
ColBERT late interaction	10-25 ms	Medium (N vectors per doc)	Fine-grained token MaxSim	High-precision candidate reranking
Cross-encoder model	50-200 ms	None (re-evaluates text)	Full cross-attention	Final top-10 scoring and reranking