Yesterday I had the opportunity to give a lecture on Introduction to Vector Databases at the Brandenburger Linux User Group. It was an exciting exchange, and for all those who couldn’t be there (or just wanted to review the content in peace), I am publishing my presentation slides here.
The topic of Artificial Intelligence is becoming more prevalent, and with it the underlying database technologies are becoming increasingly important. But what exactly makes a vector database?
Here are the most important takeaways from my lecture:
- What are vectors and embeddings? A vector is fundamentally an ordered list of numbers. In the AI context, we usually talk about high-dimensional vectors (e.g., with 768 or 1536 dimensions), where each entry represents a learned property. Models convert data such as text, images, or audio into these so-called vector embeddings.
- Vector Databases vs. Relational Databases: While classical databases search for exact keyword matches, vector databases are designed for similarity search. They find objects that are semantically similar based on distance measures.
- The Evolution of Language Understanding: We looked at how we go from text to vectors. Starting from simple methods like Bag of Words (which unfortunately don’t understand context or word order) through Word2Vec to contextual embeddings of the latest generation (State of the Art) like BERT. The latter use Transformer architectures to build genuine language understanding by assigning a different vector to the same word depending on the sentence context.
- Architectures in Practice (RAG): A central use case is Retrieval Augmented Generation (RAG). Here, we use vector databases to find the most relevant text passages for a prompt and pass these as context to a Large Language Model (LLM).
- My Practical Projects: Theory is good, practice is better! That’s why I also presented two of my own projects:
- Dinofind: An image search engine based on 30,000 Flickr images, which were converted into vectors using the DINOv2 model and stored in a Qdrant vector database.
- Malwareuniverse: A project for malware analysis.
Finally, there was a brief outlook on topics like benchmarking embedding models on own data.
Download Slides and Materials
A big thank you to everyone who was there yesterday!
If you have any questions or want to discuss the topics, feel free to email me.
