Skip to content
ReviewRamble

Qdrant

Rust vector database with fast filtered search, hybrid sparse-dense retrieval and quantization, self-hosted or managed.

Reviewed by Sheela · Updated

View on GitHub Open source

Our rating

4.7/5

Filtered vector search done properly makes Qdrant the open-source database to beat for production RAG, as long as someone locks down a server that ships open.

What it is

Qdrant is a vector database written in Rust. It stores embeddings with payloads (metadata) and applies filters during index search, so a query restricted to one customer or date range still finds the right results quickly.

It supports hybrid sparse and dense search, multi-vector search and several quantisation methods. Version 1.19 (August 2026) made 4-bit TurboQuant a primary storage type, and v1.20.0 (22 October 2026) added 8- and 16-bit TurboQuant types and BM25 scoring on the full-text index. It ships as a single binary or Docker image, and Qdrant Edge runs it on-device.

The core is Apache 2.0. Qdrant Cloud is the managed service, with a free 1 GB cluster; the company announced a $50M Series B led by AVP on 12 March 2026.

Our review

Qdrant is the strongest dedicated open-source vector database for workloads where filters matter, but its self-hosted defaults assume you will secure it yourself. The repository has about 35,200 GitHub stars, v1.20.0 shipped on 22 October 2026, and the March 2026 Series B eased most worries about the company's staying power.

Filtering is the technical reason to choose it. Because payload filters are applied while traversing the HNSW graph rather than afterwards, multi-tenant and metadata-heavy queries keep their recall, a problem pgvector users solve with iterative scans and partitions. Sizing is predictable too: Qdrant's own formula puts one million 1,024-dimension vectors at about 5.7 GB of RAM before quantisation, which it says cuts memory by a factor of 4 to 32.

The 2026 security record is the caveat. The docs say plainly that a self-hosted instance binds to all interfaces with no authentication and no TLS. In clustered mode the internal gRPC port, 6335, accepted unauthenticated requests until v1.20.0, so anyone who could reach it could add a rogue peer and receive replicated data; Qdrant rated that critical and fixed it by enforcing internal auth whenever an API key is set. Earlier, CVE-2026-25628 allowed arbitrary file writes through the /logger endpoint. Self-hosted instances also send anonymised usage statistics by default.

Milvus is built from the start as a distributed system for billions of vectors across many nodes, and pgvector remains the simpler answer for teams already on Postgres with a few million vectors. Qdrant sits between them: a separate service worth running once filtered recall and latency start to matter, provided someone owns its network exposure.

Where it earns its place

  • A platform engineer at a B2B SaaS company with thousands of customer workspaces

    Keep each tenant's documents separate inside one RAG index

    Store a tenant ID in the payload and filter on it; Qdrant applies the filter during search, so small tenants are not starved of results. Budget roughly 5.7 GB of RAM per million 1,024-dimension vectors before quantisation, per Qdrant's capacity-planning guide.

  • A search engineer on a developer-tools product

    Run keyword and semantic code search in one system

    v1.20.0 adds BM25 scoring directly on the full-text index and a code tokeniser alongside sparse and dense vectors, so hybrid queries no longer need a second search engine. The trade is a younger full-text feature set than Elasticsearch's.

  • An engineering manager at a 30-person startup

    Decide between Qdrant Cloud and self-hosting

    The free tier is one node with 0.5 vCPU, 1 GB RAM and 4 GB disk; Standard bills vCPU, memory and storage hourly with a 99.5% SLA, while SSO and a 99.9% SLA need Premium, which has an unpublished minimum spend. Self-hosting is free but leaves API keys, TLS and patching to the team.

Before you commit

Setup
docker run qdrant/qdrant or the single binary gives a REST API on port 6333 and gRPC on 6334 within minutes, with official client libraries for the common languages. Production setups must add an API key and TLS, both off by default.
Integrations
LangChain, LlamaIndex and Haystack ship Qdrant integrations; snapshots go to S3, GCS or Azure storage, and Qdrant Edge embeds the engine on devices. Qdrant Cloud adds free inference on selected embedding models.
Data and privacy
Self-hosted instances send anonymised usage statistics by default (system info, timings, crash backtraces; never stored data or collection names); QDRANT__TELEMETRY_DISABLED=true stops it. API keys, read-only keys and JWT RBAC are opt-in. 2026 advisories: arbitrary file write via /logger (CVE-2026-25628) and two internal gRPC auth flaws fixed by v1.20.0.
Cost at scale
Self-hosting is free; plan RAM with Qdrant's formula (vectors x dimensions x 4 bytes x 1.5), so 10 million 1,024-dimension vectors need about 57 GB before quantisation. Qdrant Cloud Standard is billed hourly with prices only in its calculator; Hybrid and Private Cloud are quoted by sales.
Limits
About 800 open issues and pull requests. Rolling upgrades must go one minor version at a time and the internal port still has no message-size limit in v1.20.0.

What's good

  • Filters applied during graph search keep recall high
  • Hybrid sparse, dense and BM25 retrieval in one engine
  • Single binary with a published memory-sizing formula
  • Free 1 GB managed cluster for prototypes

What's not

  • Self-hosted defaults: no auth, no TLS, all interfaces
  • Critical cluster-auth flaw fixed only in v1.20.0
  • Cloud prices only available through a calculator

Best for

Teams running production RAG or search that need fast filtered vector queries.

Alternatives

pgvector

RAG & dataC23k stars

4.6/5

Postgres extension that adds vector columns and HNSW or IVFFlat indexes, so embeddings live next to the rows they describe.

Milvus

RAG & dataGo46k stars

4.4/5

Distributed vector database for billion-scale search, now able to query vectors in Parquet, Lance and Iceberg files where they sit.

Chroma

RAG & dataRust30k stars

4.1/5

Open-source vector database that runs embedded in Python or as a server, popular for prototypes and small RAG apps.

LlamaIndex

RAG & dataPython53k stars

4.3/5

Python framework for connecting LLMs to your documents, with ingestion, indexing, retrieval and event-driven agent Workflows.

Haystack

RAG & dataPython27k stars

4.2/5

deepset's open-source Python framework for production RAG pipelines and agents, built from explicit, typed components.

Qdrant: common questions

Is Qdrant worth it?
Filtered vector search done properly makes Qdrant the open-source database to beat for production RAG, as long as someone locks down a server that ships open. We rate it 4.7 out of 5.
How much does Qdrant cost?
Free and open source under the Apache 2.0 licence. Qdrant Cloud has a free 1 GB cluster, then usage-based billing; Premium, Hybrid and Private Cloud are priced by sales.
What are the downsides of Qdrant?
Self-hosted defaults: no auth, no TLS, all interfaces; critical cluster-auth flaw fixed only in v1.20.0; cloud prices only available through a calculator.
Who is Qdrant best for?
Teams running production RAG or search that need fast filtered vector queries.
How much does Qdrant cost for a team or in production?
Self-hosting is free; plan RAM with Qdrant's formula (vectors x dimensions x 4 bytes x 1.5), so 10 million 1,024-dimension vectors need about 57 GB before quantisation. Qdrant Cloud Standard is billed hourly with prices only in its calculator; Hybrid and Private Cloud are quoted by sales.
Is Qdrant safe for company data?
Self-hosted instances send anonymised usage statistics by default (system info, timings, crash backtraces; never stored data or collection names); QDRANT__TELEMETRY_DISABLED=true stops it. API keys, read-only keys and JWT RBAC are opt-in. 2026 advisories: arbitrary file write via /logger (CVE-2026-25628) and two internal gRPC auth flaws fixed by v1.20.0.
What are the best Qdrant alternatives?
The closest alternatives we have reviewed are pgvector, Milvus, Chroma, LlamaIndex and Haystack.

Answers checked Oct 23, 2026.

Make Qdrant? Show this rating on your siteQdrant rated 4.7 out of 5 on ReviewRamble