How much RAM does Qdrant need?

Qdrant itself idles around 100–200 MB. The practical limit is your index size: 1 million 1536-dimensional float32 vectors take about 6 GB unquantised. With scalar quantisation (int8) that drops to ~1.5 GB. A 4 GB VPS is comfortable for most small-to-medium projects.

Which embedding models and frameworks work with Qdrant?

Any model that produces fixed-size float vectors works. Popular choices self-hosted via Ollama: `mxbai-embed-large` (1024 dims), `nomic-embed-text` (768 dims). Via API: OpenAI `text-embedding-3-small` (1536 dims), Cohere Embed. Frameworks with native connectors: LangChain, LlamaIndex, Haystack, semantic-kernel, and the official Python/TypeScript/Go/Rust SDKs.

Is Qdrant a good alternative to Pinecone or Weaviate?

Qdrant competes directly with Pinecone (managed) and Weaviate (OSS). Qdrant's advantages: written in Rust for lower memory overhead, the best payload-filter + vector-search combination in benchmarks, and Apache 2.0 licence with no phoning home. Trade-off: you manage the infrastructure yourself, which is exactly what a ServOrbit VPS gives you.

Can I run Qdrant alongside Ollama on the same VPS?

Yes. Both are single-container deployments with no port conflict (Ollama on 11434, Qdrant on 6333/6334). A 4 GB RAM VPS comfortably hosts both with a 7B Q4 model loaded. This is the recommended starting point for a fully self-contained, air-gapped RAG stack.

Build, host and run AI solutions.

Qdrant

Self-hosted vector database for semantic search and RAG — the missing infrastructure layer for your LLM stack.

2 GB RAM 1 vCPU Port 6333 Available

Source code

Tech stack

DockerRust

Minimum RAM2 GB

Minimum CPU1 vCPU

Default port6333

Compatible OSubuntu-24.04

Qdrant is an open-source, high-performance vector database written in Rust. It stores vector embeddings — the numerical representations produced by LLMs and embedding models — and retrieves the most similar ones in milliseconds, even across hundreds of millions of vectors.

Deployed on a ServOrbit VPS, Qdrant becomes the memory layer of your AI stack: pair it with Ollama for local inference, LiteLLM for a unified API gateway, and your own application code to build production-grade RAG pipelines, semantic search engines, recommendation systems or AI chatbots that actually remember context — all on infrastructure you control, at a fixed monthly cost.

Key features

Sub-millisecond approximate nearest-neighbour search over millions of vectors

Payload filters — combine vector similarity with structured filters (category, date, score) in a single query

HNSW indexing + scalar and product quantisation for fast search on small RAM budgets

Collections and named vectors — store multiple embedding types per document in one collection

REST and gRPC APIs — every popular LLM framework (LangChain, LlamaIndex, Haystack) has a native Qdrant connector

Apache 2.0 — fully auditable, no telemetry, air-gap capable

When to use this solution?

1

RAG pipeline memory

Index your documentation, knowledge base or product catalogue as embeddings and retrieve the most relevant chunks before each LLM call. Pair with Ollama on the same VPS for a fully air-gapped, zero-per-token RAG stack.

2

Semantic search

Replace keyword search with meaning-based retrieval. A query for 'fast payment processing' surfaces results about 'instant transactions' even when those exact words never appear — because Qdrant compares intent, not tokens.

3

Recommendation engine

Represent products, articles or users as embedding vectors and find the nearest neighbours. 'Users who liked this' and 'similar products' become a single nearest-neighbour query with optional payload filters for price or category.

Deploy Qdrant on your VPS

Guide optimized for ServOrbit Cloud VPS.

01

Order a VPS

Pick a ServOrbit VPS with at least 2 GB RAM (512 MB covers Qdrant itself; the rest is headroom for your embedding model or co-located services). Ubuntu 22.04, Docker pre-installed by the provisioner.

02

Deploy Qdrant

The provisioner runs: docker run -d -p 6333:6333 -p 6334:6334 -v qdrant_storage:/qdrant/storage --name qdrant --restart always qdrant/qdrant:latest. The REST API is immediately available on port 6333; gRPC on 6334.

03

Create a collection and index vectors

Create a collection with the right vector size for your model (e.g. 1536 for OpenAI text-embedding-3-small, 4096 for mxbai-embed-large via Ollama): curl -X PUT http://:6333/collections/my_docs -H 'Content-Type: application/json' -d '{"vectors":{"size":1536,"distance":"Cosine"}}'. Then upsert your embeddings via the REST API or any Qdrant SDK.

04

Query from your application

Install the Qdrant client (pip install qdrant-client or npm install @qdrant/js-client-rest) and query: pass the embedding of the user's question and get back the top-k most similar document chunks. Feed those chunks as context to your LLM for a grounded, hallucination-resistant answer.

05

Secure the instance

Set QDRANT__SERVICE__API_KEY as an environment variable (restart the container with -e QDRANT__SERVICE__API_KEY=) and put Qdrant behind a reverse proxy (Caddy or Nginx) for HTTPS. Never expose port 6333 directly to the internet without authentication.

Frequently asked questions

A vector database stores embeddings — the numerical representations produced by LLMs and embedding models. RAG (Retrieval-Augmented Generation) works by retrieving the document chunks most similar to a user's question before calling the LLM. A vector database like Qdrant does that retrieval in milliseconds across millions of chunks, which a relational database or full-text index cannot match.