Key takeaways
- Qdrant is an open-source vector database written in Rust, built for similarity search with rich filtering.
- It stores points made of an ID, one or more vectors and a JSON payload, indexed with HNSW.
- Qdrant was founded in 2021 by Andre Zayarni and Andrey Vasnetsov; the company is based in Germany.
- Choose Elasticsearch when full-text search leads; choose Qdrant or Milvus for vector-first workloads.
- Run it locally with one Docker command, or use the Python client's local mode without a server.
Qdrant is an open-source vector database, written in Rust, for storing embeddings and searching them by similarity with fast metadata filtering. It runs as a single service you can self-host or use through Qdrant Cloud, and is a popular choice for RAG, semantic search and recommendations. I build RAG systems such as RAG.NextUpgrad with hybrid retrieval, so this page explains what Qdrant is, how it works and stores data, who made it and when, how it compares with Elasticsearch and Milvus, images, and installation.
What is Qdrant DB?
Qdrant is a vector database and similarity search engine. You store vectors, such as text or image embeddings, with JSON payloads, then query for the nearest vectors to a query vector, optionally filtered by payload fields such as category, date or user.
Also asked as: how qdrant db works
It supports dense vectors, sparse vectors for keyword-style hybrid search, multiple named vectors per record, and multi-vector representations. It offers REST and gRPC APIs with client libraries for Python, JavaScript, Rust, Go, Java and .NET. It is licensed under Apache 2.0. Typical use cases include RAG retrieval, semantic product search, recommendations, duplicate detection and anomaly detection.
How does Qdrant work?
Qdrant organises data into collections of points. Each point has an ID, vectors and a payload. It builds an HNSW graph index over vectors for fast approximate search, and payload indexes for filtering, applying filters during graph traversal so results stay accurate.
When you query, Qdrant walks the HNSW graph towards the query vector while skipping points that fail the filter, then returns the top matches with scores and payloads. It supports cosine, dot product, Euclidean and Manhattan distances. Quantisation options compress vectors to cut memory, with rescoring on original vectors for accuracy. For hybrid search, it can combine dense and sparse results with fusion in a single query. Distributed mode shards and replicates collections across nodes.
How does Qdrant store data?
Qdrant stores each collection in segments containing vectors, payloads and indexes. Writes go first to a write-ahead log for durability, then into segments, which are optimised and merged in the background.
You can keep vectors in memory for speed or on disk with memory-mapped files to handle collections larger than RAM, and similarly choose where payloads live. Snapshots let you back up and restore collections. When self-hosting, mount the storage directory on persistent disk, otherwise data disappears when a container is removed. Quantised vectors can stay in memory while full-precision originals stay on disk, a common way to balance speed and cost.
Who made Qdrant?
Qdrant was created by Andre Zayarni, the company's chief executive, and Andrey Vasnetsov, its chief technology officer, who began it as an open-source project. Qdrant the company, based in Germany, develops the database and runs Qdrant Cloud.
The project is developed in the open on GitHub, with contributions from the community. The company is venture-backed and offers managed cloud, hybrid cloud and enterprise options alongside the free open-source version. For users, this means a commercially supported product with an open-source core, so you can self-host without licence fees or move to managed hosting later.
When was Qdrant founded?
Qdrant the company was founded in 2021; the open-source project started a little earlier, around 2020, with early releases on GitHub. It grew quickly as RAG made vector search a mainstream need after 2022.
Since then, it has added features such as sparse vectors, hybrid queries, multi-vector support, quantisation, distributed deployment and a managed cloud. Vector databases as a category are young, so features change rapidly; check the release notes and documentation for current capabilities rather than relying on older comparisons.
Qdrant vs Elasticsearch: which should I use?
Use Elasticsearch when full-text search, aggregations and log analytics are central and vector search is an addition. Use Qdrant when vector similarity with filtering is the core workload and you want a lighter, specialised engine.
Elasticsearch has mature BM25 search, analysers for many languages, aggregations and a huge ecosystem, plus dense-vector kNN search. It is heavier to run and tune. Qdrant is purpose-built for vectors, efficient with filtered search and quantisation, and simpler to operate, with sparse vectors for keyword-style signals. Many RAG teams pair a keyword engine with a vector engine, or use one that does both well enough; measure on your own queries before deciding.
Qdrant vs Milvus: what is the difference?
Both are open-source vector databases. Milvus is a distributed, cloud-native system built for very large scale, with separate components and GPU index options. Qdrant is a single Rust service that is simpler to run, with optional distributed mode.
Milvus suits teams expecting billions of vectors and willing to operate a more complex architecture, or using Zilliz Cloud. Qdrant suits teams wanting quick setup, strong filtering and efficient resource use, scaling out when needed, or using Qdrant Cloud. Performance comparisons depend heavily on dataset, recall target and configuration, so run your own benchmark with your vectors and filters rather than trusting vendor charts.
Can Qdrant store images?
Qdrant stores image embeddings, not image files. You compute vectors with an image model such as CLIP, store them in Qdrant with a payload containing the image's URL or storage path, and keep the actual images in object storage or a file system.
Payloads can technically hold encoded image data, but it bloats the database and slows queries, so avoid it. This pattern enables image search by text, finding similar images and duplicate detection. It also helps privacy: FaceVision follows the same principle, storing face embeddings and never the images themselves. If you need image and text search together, store multiple named vectors per point.
How do I install Qdrant locally?
The simplest way is Docker: pull the image and run it with ports 6333 for REST and 6334 for gRPC, mounting a folder for storage. The dashboard is then available on port 6333 in your browser.
Also asked as: how to install qdrant locally
docker pull qdrant/qdrant
docker run -p 6333:6333 -p 6334:6334 \
-v "$(pwd)/qdrant_storage:/qdrant/storage" qdrant/qdrant
Then pip install qdrant-client and connect to localhost on port 6333. For quick experiments without any server, the Python client has a local mode that runs in memory or saves to a folder. Enable an API key before exposing Qdrant beyond your machine.
How do I install Qdrant on Windows?
On Windows, install Docker Desktop with the WSL 2 backend, then run the same Docker command as on other systems, using a Windows path or named volume for storage. Alternatively, use the Python client's local mode for development without running a server.
Also asked as: how to install qdrant in windows
In PowerShell, mount storage with a full path such as -v ${PWD}/qdrant_storage:/qdrant/storage, or use a named Docker volume to avoid path issues. Qdrant also publishes release binaries; check the GitHub releases page for current Windows support. For production, run Qdrant on Linux servers or a managed cloud, and treat Windows setups as development environments.
Sources
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
