A key element in building an RAG(Retrieval Augmented Generation) is the usage of a vector database. The list of vector databases is growing every day and even SQL Server now supports vectors. In this post I focus on Qdrant (pronounced as "quadrant") a vector similarity search engine and database written in Rust. It's specifically designed to handle vector embeddings and payload, making it perfect for modern machine learning applications.
The goal of this post is not to make you a Qdrant expert(neither am I) but to provide you with enough information to start using it correctly.
What is Qdrant?
Installing Qdrant
Setting up Qdrant is straightforward. You can run it using Docker:
docker pull qdrant/qdrant docker run -p 6333:6333 qdrant/qdrant
Or download and install it directly on your system using a release found here:
https://github.com/qdrant/qdrant/releases/
After deploying Qdrant locally, you can access the Web UI using the following URI: http://localhost:6333/dashboard
We come back to this Web UI later, but let me first explain the basic building blocks in Qdrant.
The basic building blocks
The main building blocks you need to understand are:
- Collection: A named set of points that you can use for your search.
- Point: A record which consists of a vector and an optional payload.
- Distance Metrics: The method used to measure similarities among vectors.
Collections
A collection is a named set of points (vectors with a payload) among which you can search. You can compare a collection with a table in relational database. The vector of each point within the same collection must have the same dimensionality and be compared by a single metric.
Qdrant supports two modes for vector storage. The default mode is single unnamed vector. In this case a collection may only contain a single vector and it will be unnamed in the storage model in Qdrant. A second mode is named vector. By using this mode it is possible to have multiple vectors in a single point, each of which can have their own dimensionality and metric requirements.
Remark: Semantic Kernel can handle both unnamed and named vector modes.
Points
The points are the central entity that Qdrant operates with and they consist of a vector and an optional id and payload.- id: a unique identifier for your vectors.
- Vector: a high-dimensional representation of data, for example, an image, a sound, a document, a video, etc.
- Payload: A payload is a JSON object with additional data you can add to a vector.
Every vector has a specific size that is defined at the collection level. The size typically depends on the used embedding engine as different engines generate vectors in different sizes.
Remark: Qdrant only supports only ulong or guid id’s. Choosing a different id type in Semantic Kernel will result in errors.
Distance Metrics
Distance metrics are used to measure similarities among vectors and they must be selected at the same time you are creating a collection. The choice of metric depends on the way the vectors were obtained and, in particular, on the neural network that will be used to encode new queries.
There are a lot of distance metrics, but Qdrant supports only some of the most popular ones:
- CosineSimilarity
- DotProductSimilarity
- EuclideanDistance
- ManhattanDistance
Here is a short explanation of each of them :
- Cosine Similarity - Cosine similarity is a way to measure how similar two vectors are. To simplify, it reflects whether the vectors have the same direction (similar) or are poles apart. Cosine similarity is often used with text representations to compare how similar two documents or sentences are to each other.
- Dot Product - The dot product similarity metric is another way of measuring how similar two vectors are. Unlike cosine similarity, it also considers the length of the vectors. This might be important when, for example, vector representations of your documents are built based on the term (word) frequencies.
- Euclidean Distance - Euclidean distance is a way to measure the distance between two points in space, similar to how we measure the distance between two places on a map.
- Manhattan Distance – Manhattan distance is similar to the Euclidean in the sense that it also is used to measure the distance between two points in space, the difference is that it is using a grid-like path by summing the absolute differences of their corresponding coordinates.
Remark: When using Semantic Kernel and you don’t explicitly specify a distance metric/function, Cosine Similarity will be used by default.
Creating and searching our first data in Qdrant
Let’s now apply the information above to create our first collection in Qdrant. Therefore let’s go back to the Web UI available at http://localhost:6333/dashboard.
Click on the Quickstart button to open the Quickstart tutorial and let us walk through the steps.
First we need to create a collection to store our data. In this example a collection called star_charts will be created. This collection will store location data where each location will be represented by a vector of four dimensions, and we'll use the Dot product as the distance metric for similarity search.
The Web UI has a built-in console allowing us to run commands directly in the web user interface. We’ll use this to create our collection with the specified vector size and distance function:
We can see the created collection by clicking on the Collections icon in the left toolbar:
Now we can upload some data. This can be done directly as shown in the tutorial:
Another option is to upload a snapshot:
Or import a dataset (Qdrant comes out-of-the box with some example datasets):
We can browse the stored data by clicking on the specific Collection:
And even have a graph view to visualize the relative distance:
Last step is to run our first search. We specify a (calculated) vector, the number of vectors we want to get back and if the payload should be returned or not:
That’s it! With this information you know enough to start using Qdrant.
More information
An Introduction to Vector Databases – Qdrant
Exploring Microsoft.Extensions.VectorData with Qdrant and Azure AI Search - .NET Blog
Using the Semantic Kernel Qdrant Vector Store connector (Preview) | Microsoft Learn