Qdrant

What it is for

Qdrant is a vector search engine: create a collection, upsert points with a payload, then query_points with filters. Use it when the job is store and search embeddings you already made, not when you need the embedder itself.

What makes it unique: the cover is the official qdrant/qdrant GitHub file list — src/, docs/, lib/, Apache-2.0 in the About pane — not a marketing orb.

Key features

• Collections + search — create a collection (size + distance), upsert points, query_points (Local Quickstart; Qdrant Essentials). One engine; HNSW / quantization sit on this index, not a second product.
• Payload filters — attach JSON (city: London) and query_filter so a nearest-neighbor also matches a field (Quickstart “Add a filter”). A predicate, not a vector.
• Hybrid / sparse — dense vectors plus BM25 / SPLADE-style sparse in the same collection (Cloud “SPLADE, BM25”). A second signal, not a second database.
• Cloud clusters — Free forever (0.5 vCPU / 1 GB / 4 GB) or Standard dedicated nodes billed hourly (pricing; Create a Cluster). A host, not the engine.
• Cloud Inference — generate text/image embeddings inside the cluster (MiniLM, SPLADE, BM25, Mixedbread, CLIP) so you do not call a second vendor (Cloud FAQ). An embed pipe on the store, not Voyage.

Docker / QdrantClient(url=...) and Qdrant Edge (in-process) are install / SDK. vCPU / RAM / disk hours (and paid inference tokens) are the meter. GPU SKUs and cloud regions are SKU.

Who it's for

Best fit: RAG teams who will run OSS or Cloud Free for $0 (1 GB / 4 GB, ~1M × 768-d vectors) and only open Standard when they need dedicated RAM and a 99.5% SLA.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from qdrant.tech/pricing, Create a Cluster, and the Apache-2.0 LICENSE.

• Software (OSS) — Apache 2.0. License fee $0. You host Docker / Kubernetes / Edge.
• Cloud Free — $0 forever. No card. 1 node: 0.5 vCPU / 1 GB RAM / 4 GB disk (~1M × 768-d). Suspends after 1 week idle; deleted after 4 weeks if not reactivated. Manual snapshots only.
• Cloud Standard — usage-based (vCPU + RAM + disk + backups + paid inference tokens), billed hourly. Dedicated resources; HA; 99.5% SLA; free tokens on paid inference models (up to 5 million / model / month on paid clusters; BM25 uncapped). No public seat dollar on the cards we opened — use the sizing calculator; billing docs use a worked $85/mo example, which is not a plan card.
• Cloud Premium — minimum spend. SSO; private VPC links; 99.9% (99.95% multi-AZ) SLA.
• Hybrid / Private Cloud — you bring infra or on-prem; quoted.

Do not invent a $ / GB-RAM row we did not copy from the live calculator.

Limitations / costs

• Free Cloud suspends after 1 idle week. Reactivate or you lose the cluster at 4 weeks.
• Standard has no printed seat. Confirm the calculator before you buy.
• A store, not an embed API. For specialist voyage-4 vectors, see Voyage AI.

Alternatives

• Voyage AI — hosted embeddings/rerank when Qdrant Inference’s MiniLM/CLIP set is not the model you want.
• Chroma — another local vector store when you want a Python-first library, not a Rust engine.
• LlamaIndex — indexes + LlamaParse when you want the retrieval framework more than the engine.

Official tutorial

Official Qdrant Vector Search channel video: Qdrant Essentials | Building Simple Vector Search in Qdrant (oEmbed author: Qdrant Vector Search). First-hour Cloud collection + upsert + search.

Written path: Local Quickstart plus Pricing. Channel also has Semantic Search for Beginners (same official author).

Website

https://qdrant.tech

Repository

https://github.com/qdrant/qdrant

Source

Catalog row curated for AI Tools Nav. Cover is the official qdrant/qdrant GitHub file list (Apache-2.0 + src/docs). Dollars checked on qdrant.tech/pricing (OSS $0; Cloud Free $0 / 0.5 vCPU / 1 GB / 4 GB; Standard usage-based, no seat card). Tutorial oEmbed author: Qdrant Vector Search.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.