Qdrant

What is Qdrant?

Qdrant is a vector database — a store that keeps text, images, or audio as numeric embeddings and answers “what is most similar to this?” instead of “what contains this word?”. That question is the foundation of retrieval-augmented generation, where an AI application looks up the handful of passages from your own documents that actually relate to a user’s question and feeds them to a language model, so the answer is grounded in your content rather than invented. The same lookup powers semantic search, agent memory, and recommendations. Qdrant is written in Rust, indexes with HNSW, and applies metadata filters during index traversal rather than after retrieval — which avoids the common failure where a restrictive filter leaves a query with almost nothing to return.

The engine is open source under an unmodified Apache 2.0 license and can be self-hosted for free with no usage caps or license keys; sharding, replication, hybrid search, and quantization all ship in the free release rather than being held back for a paid edition. Qdrant Cloud is the managed alternative, running clusters on AWS, Azure, or GCP with a free-forever starter cluster and usage-based billing above it, while Hybrid Cloud runs managed clusters on infrastructure you own. The proprietary part is the operational layer — provisioning, scaling, multi-AZ topology, SSO, private networking — not the search engine itself. The project passed 33,000 GitHub stars, shipped version 1.18.3 in July 2026, and the company raised a 50 million dollar Series B in March 2026.

Who is it for?

Qdrant suits engineering teams building retrieval features who care about where their data sits and what happens if pricing changes. Because the engine is the same whether you run it yourself or buy the managed service, you can prototype on the free cloud tier, move to Standard when traffic justifies it, and still keep self-hosting as a real fallback rather than a theoretical one. It is developer infrastructure: you will be writing code against an API, not clicking through a builder.

• AI application developers building RAG pipelines who want filtered, hybrid retrieval and are comfortable choosing an embedding model and managing a collection schema.
• Teams with data-residency or compliance constraints who need the database inside their own network or an air-gapped environment, via self-hosting, Hybrid Cloud, or Private Cloud.
• Cost-sensitive startups whose query volume makes per-query managed pricing painful, and who would rather pay for a server than a metered retrieval API — see Pinecone for the fully managed contrast.
• Platform and infrastructure engineers standing up shared retrieval for several internal applications, who need sharding, replication, and quantization to control memory cost at scale.
• Agent builders who need durable semantic memory and want to query it through the official MCP server or the Python, JavaScript, Rust, Go, Java, and .NET clients.