Qdrant

Open-source vector database and similarity search engine. Specializes in high-performance ANN search for production ML workloads. Competes with Pinecone, Weaviate, Milvus.

Notable Contributions

TurboQuant (Qdrant 1.18) — rotation-based vector quantization; extends Google Research’s TurboQuant paper with anisotropy compensation (per-coordinate calibration), length renormalization from RaBitQ, full L2/dot/cosine support, and SIMD kernels. Results: 8× compression at SQ-level recall; +10–20 pp recall over BQ at 16×/32× storage.

Scalar Quantization — int8 per coordinate, 4× compression, near-lossless.

Binary Quantization — 1-bit/2-bit storage, 16×–32× compression.

Index-native relevance feedback (Qdrant 1.17, Feb 2026) — a RelevanceFeedbackQuery API that folds model-generated Relevance Feedback into HNSW traversal: the hop-selection function becomes a mix of query similarity and feedback scores, so the feedback steers the walk through the whole collection rather than reranking a retrieved top-k. Three parameters, fitted once per (feedback model, collection, retriever) via qdrant-relevance-feedback; effectively Knowledge Distillation of a reranker into the index. Positioned as the first vector-index-native relevance feedback API — possible only because Qdrant owns its index.

Sparse neural retrieval — Qdrant supports BM25 and SPLADE as sparse vectors over inverted indexes, making lexical and learned-sparse retrieval first-class in a vector engine. Original research includes miniCOIL (BM25 extended with a small semantic component, with BM25 fallback for out-of-vocabulary terms) and a five-part study of SPLADE domain fine-tuning, Fine-Tuning Sparse Embeddings for E-Commerce Search, packaged as qdrant-sparse-finetune.

People

• Ivan Pleshkov
• Jonas Schulz
• Evgeniya Sukhodolskaya — developer advocate; miniCOIL, sparse neural retrieval
• Thierry Damiba — developer relations; sparse fine-tuning research

Articles

• TurboQuant in Qdrant

• Choosing a Vector Database for ANN Search at Reddit — qualitative score 292 vs Milvus 281; better raw latency but lost on Go ecosystem fit and automatic rebalancing

• Fine-Tuning Sparse Embeddings for E-Commerce Search — Thierry Damiba

• Relevance Feedback in Qdrant — Evgeniya Sukhodolskaya; the index-native feedback formula, training and BEIR results

• Relevance Feedback in Informational Retrieval — Evgeniya Sukhodolskaya; the survey arguing the problem can only be solved inside the engine

Videos

• Evgeniya Sukhodolskaya - Fine-Tuning Sparse Neural Retrievers for E-Commerce — MICES 2026
• Evgeniya Sukhodolskaya - Relevance Feedback Inside the Search Engine — Berlin Buzzwords 2026

Tools

• Qdrant Vector DB
• qdrant-sparse-finetune — open-source SPLADE fine-tuning framework
• qdrant-relevance-feedback — fits the relevance feedback scoring parameters