Qdrant

Qdrant is an open-source vector similarity search engine and database built in Rust, created by a Berlin-based team led by CEO Andre Zayarni and CTO Andrey Vasnetsov. First released in 2021 and licensed under Apache 2.0, it is purpose-built to handle one specific problem well: storing, indexing, and querying high-dimensional embedding vectors at production scale. It is not a general-purpose database. Vasnetsov has described the philosophy directly: "scalability and performance is much more important than transactional consistency, so it should be treated as a search engine rather than database." That focus shows in the implementation.

Core capabilities include dense vector search (HNSW-based), sparse vector support for keyword-style retrieval, and a Query API that combines both in multi-stage hybrid search on the server side. Payload filtering lets you combine vector similarity with structured metadata conditions in a single query. Quantization options (scalar, binary, 1.5-bit, 2-bit, asymmetric) compress vectors in memory without rebuilding indexes. Qdrant Cloud offers a free-forever single-node tier, usage-based standard pricing, and a Hybrid Cloud model where the database runs on customer infrastructure while Qdrant manages operations. The project has passed 250 million downloads and 29,000 GitHub stars as of March 2026.

What Qdrant actually does in April 2026

Qdrant accepts vector embeddings alongside arbitrary JSON payloads, builds HNSW (Hierarchical Navigable Small World) indexes optimized for approximate nearest neighbor search, and returns ranked results filtered by payload conditions. The Query API introduced in version 1.10 lets you chain retrieval stages: a first pass with quantized vectors finds candidates, a second pass rescores them with full-precision vectors, and metadata filters apply at either stage. This is done server-side, meaning your application sends one request instead of coordinating multiple round trips.

Recent versions added GPU-accelerated HNSW indexing for faster ingestion of large collections, Maximal Marginal Relevance (MMR) to balance relevance against diversity in results, ACORN algorithm for higher-quality filtered search quality, and Score-Boosting Reranking that blends vector similarity with custom business signals (recency, user tier, popularity). Qdrant Edge, released in 2025, is a lightweight embedded variant designed for AI running on devices with limited compute, including robots, point-of-sale terminals, and mobile phones. Qdrant Cloud Inference (launched July 2025) bundles dense, sparse, and image embedding generation directly with the hosted search service, removing the need for a separate embedding API call.

The self-hosted version runs as a single binary, a Docker container, or on Kubernetes. Collections, snapshots, and cluster configuration are managed via REST or gRPC API. A web-based dashboard provides basic collection browsing and search testing, though it does not support bulk operations from the UI.

Where Qdrant sits versus Weaviate and Pinecone

The three most commonly compared purpose-built vector databases are Qdrant, Weaviate, and Pinecone. The differences are architectural, not cosmetic.

Weaviate is written in Go and uses a modular architecture where vector indexes, storage backends, and search modules are separately pluggable. Go's garbage collector introduces occasional latency spikes under memory pressure. Qdrant's Rust implementation avoids GC pauses entirely, which matters for applications with strict tail latency requirements. Weaviate emphasizes an object-oriented storage model with knowledge graph capabilities and a GraphQL API, making it a better fit for applications that need relational graph traversal alongside vector similarity. Its hybrid search uses relativeScoreFusion, which preserves the original metric values from each search method. Qdrant's hybrid approach uses prefetch-and-rescore, letting you specify exactly how candidates flow between stages. Both are open-source and both support self-hosting.

Pinecone is proprietary software with no self-hosting option and no open-source license. Its architecture separates storage from compute at the infrastructure level, enabling serverless scaling to tens of billions of vectors with essentially no infrastructure management. That simplicity comes at a cost: you cannot run Pinecone on your own infrastructure, data sovereignty is not possible, and vendor lock-in is complete. For teams with GDPR constraints, regulated-industry compliance requirements, or data residency obligations, Pinecone is simply not an option. Qdrant Hybrid Cloud addresses exactly this gap: it provides a managed operations experience while keeping the data and compute inside customer-controlled infrastructure. Pinecone also has no free self-hosted tier. Qdrant's core is Apache 2.0 and free forever.

"Qdrant powers our demanding recommendation and RAG applications. We chose it for its ease of deployment and high performance at scale, and we have been consistently impressed with its results." - Srubin Sethu Madhavan, Technical Lead, HubSpot, qdrant.tech/blog/case-study-hubspot, 2024

What the production workflow actually looks like

A typical RAG pipeline with Qdrant looks like this: documents are chunked, embedded via an external model (OpenAI, Cohere, Jina, or now Qdrant Cloud Inference), and upserted into a collection with payload metadata (document ID, timestamp, source, tenant ID). At query time, the user's input is embedded and sent to Qdrant's Query API with a payload filter restricting results to the current tenant and the relevant date range. Qdrant runs the prefetch stage (fast approximate search over quantized vectors), then the rescore stage (full-precision ranking of the top candidates), and returns the final ranked list to the application, which passes it to the LLM as context.

For teams building multi-tenant SaaS on a single Qdrant cluster, Tiered Multitenancy (added 2025) allows mixing large tenants and small tenants in the same collection without per-tenant performance degradation. For recommendation systems, the workflow is similar but with behavioral embeddings: user interaction vectors stored and updated incrementally, queried at runtime against content vectors. TripAdvisor built this at 1 billion-plus review scale, reporting 2-3x revenue increases from users engaging with their AI Trip Planner compared to traditional search interfaces.

"Qdrant has been crucial for our transformation. When you're dealing with over a billion plus user-generated, multi-modal pieces of content.." - Rahul Todkar, Head of Data and AI, TripAdvisor, qdrant.tech/blog/case-study-tripadvisor, 2024

Who Qdrant is built for

Qdrant is built for ML engineers and backend engineers who need a dedicated vector search layer in a production AI system. If you are building RAG, semantic document search, personalized recommendations, image similarity search, or agentic retrieval pipelines and you need predictable low-latency performance at scale, Qdrant is a strong default choice in 2026. The Apache 2.0 license means you can run it in any environment without royalties or contractual restrictions.

Enterprise teams with data sovereignty requirements benefit most from the Hybrid Cloud model. Security and compliance teams at regulated companies (healthcare, finance, government) can deploy Qdrant inside their own AWS VPC or on-premises Kubernetes cluster while Qdrant's team handles version upgrades and operational monitoring. This is the gap that Pinecone structurally cannot fill.

The $50 million Series B raised in March 2026 signals the direction: composable vector search as permanent infrastructure, not a temporary layer to be replaced by a general-purpose database extension. Investors include AVP, Bosch Ventures, Unusual Ventures, Spark Capital, and 42CAP, with customer validation from Canva, HubSpot, Roche, OpenTable, and Argonne National Laboratory's Aurora supercomputer.

What Qdrant is not

Qdrant is not a drop-in replacement for a general-purpose database. It does not support SQL, relational joins, or transactions in the traditional sense. If your application primarily manipulates structured data and you want to add basic vector search, a Postgres extension like pgvector may be simpler and sufficient. Qdrant's strength is pure vector retrieval at scale; if you are not working with embeddings or similarity search, there is no reason to add it to your stack.

Qdrant is not suitable for teams that want zero infrastructure responsibility. While Qdrant Cloud simplifies operations significantly, self-hosted production deployments require engineering investment: cluster sizing, payload index planning, quantization strategy, backup configuration, and monitoring setup. Pinecone genuinely abstracts all of this away, at the cost of flexibility and data control.

Qdrant's dashboard UI is functional but minimal. You cannot bulk-delete collections by pattern, multi-select collections for operations, or visualize embedding spaces without external tooling. Teams that rely on database GUIs for day-to-day operations will find it limiting. The API and CLI are the primary interfaces, which suits engineers but creates friction for data teams without coding backgrounds.

Filtering performance can degrade at extreme scale. A documented GitHub issue from September 2025 showed RPS dropping from 362 to 268 under range filter queries across 260 million points. This is a known tradeoff and Qdrant has published tuning guidance, but it requires deliberate payload index design from the start, not as an afterthought.