Qdrant Vector DB

Open-source vector database and similarity search engine optimized for high-performance ANN retrieval in production ML workloads. Maintained by Qdrant (the company).

• Website: https://qdrant.tech/
• GitHub: https://github.com/qdrant/qdrant

---

What It Does

Qdrant stores dense vectors alongside payload (structured metadata) and supports filtered ANN search — combining vector similarity with attribute predicates efficiently via graph-integrated filtering.

Key capabilities:

• ANN search — HNSW-based index; high recall/speed tradeoff
• Filtered search — predicates integrated into graph traversal (not post-filter), avoiding recall loss on selective filters
• Quantization — Scalar Quantization (int8, 4×), Binary Quantization (1-bit, 32×), TurboQuant (rotation-based, 8×–32×, Qdrant 1.18+)
• Sparse + dense — supports sparse vectors alongside dense for hybrid retrieval
• Hybrid fusion — RRF (with configurable k since v1.16 and per-prefetch weights since v1.17) or DBSF (since v1.11), combining prefetch results at query time
• Multivectors / late interaction — native multivector_config with MultiVectorComparator.MAX_SIM for ColBERT/ColPali rerank; see Late Interaction in Qdrant
• Payload filtering — rich JSON payload per vector with indexed field filtering
• Collections — top-level namespace; each collection has its own vector config and index
• Memory tiers — unified pinned/cached/cold setting per component (v1.19+), replacing the on_disk, always_ram, and on_disk_payload flags
• Per-tenant IDF — BM25 corpus statistics scoped to a tenant inside a shared collection (v1.19+); see Multi-Tenancy in Search

Quantization Options

Method	Compression	Notes
Scalar Quantization (int8)	4×	Near-lossless; default recommended
Binary Quantization	16×–32×	Significant recall loss without oversampling
TurboQuant (v1.18+)	8×–32×	Rotation-based; beats BQ by 9–24 pp recall
turbo4 datatype (v1.19+)	9× on disk	Stores only the 4-bit code; no originals retained, so no rescoring

Related Tools

• Weaviate Vector DB — competing vector database; native cross-encoder reranking support

• Pinecone — managed-only competitor

• FAISS — library (not a service); no filtering, no persistence

• qdrant-es-gateway — third-party Elasticsearch-compatible REST front end that uses Qdrant as a lexical engine: server-side BM25 sparse vectors, payload filters and facets; its documented surface has no kNN or dense-vector query

• FastEmbed — Qdrant’s embedding library; computes MUVERA encodings client-side

Related Concepts

• HNSW — the ANN index Qdrant uses
• Vector Quantization — compression family
• TurboQuant — Qdrant’s latest quantization (1.18)
• Dense Vector Retrieval — primary use case
• Vector Filtering — Qdrant’s predicate-in-graph approach
• Hybrid Search — combining sparse and dense retrieval
• Reciprocal Rank Fusion — configurable k/weights fusion method
• Distribution-Based Score Fusion — Qdrant’s score-preserving fusion alternative to RRF
• Late Interaction — native multivector MaxSim reranking; see Late Interaction in Qdrant
• MUVERA — fixed-dimensional approximation used as the late-interaction first stage

Articles

• TurboQuant in Qdrant

• Qdrant 1.19 - Turbo4 Datatype and Memory Tiers — turbo4 storage datatype, memory tiers, per-tenant IDF, slice filtering

• Choosing a Vector Database for ANN Search at Reddit — head-to-head vs. Milvus at 340M vectors; Qdrant showed better raw latency at RF=1 but lost on scaling and organizational fit

• How to Tune Hybrid Search in Qdrant — Dylan Couzon; RRF vs. DBSF and RRF k/weight tuning, measured across five datasets

• Hybrid Queries - Qdrant — RRF and DBSF fusion query reference

• MUVERA in Qdrant - Multi-Vector Search Course — MUVERA prefetch via FastEmbed, MaxSim rerank

People

• Ivan Pleshkov — TurboQuant implementation
• Jonas Schulz — TurboQuant co-author
• Mohamed Arbi Nsibi — 1.19 release writeup
• Dylan Couzon — hybrid search tuning guide