On Vector Databases. Q&A with Frank Liu
Q1. Vector embeddings are becoming increasingly popular in natural language processing (NLP), computer vision, and other artificial intelligence (AI) applications. Why?
Vector embeddings are the words and phrases that computers use to communicate. Unlike us, who speak various human languages like English, Japanese, French, or Spanish, computers communicate using numbers, bits, bytes, ins, floats, etc. Vector embeddings are a great way to represent knowledge and data in a way that computers can understand. In fact, they are the primary method that computers use to represent data. This is why you need vector embeddings. That’s also why a purpose-built vector database is necessary to utilize vector embeddings effectively.
Q2. This has led to the emergence of vector databases. What are the key differences between vector databases and traditional databases?
Traditional databases are designed to handle table-based or relational data and NoSQL data such as objects and documents. However, they are not optimized for vector embeddings.
In contrast, vector databases are specifically designed to handle vector embeddings from the ground up. They excel at conducting approximate nearest-neighbor searches on a very large scale. This means they can search for other vectors close to a given query vector in a high-dimensional space.
Overall, vector databases are purpose-built for vector search and are the best option for handling vectors.
Q3. How does a vector database work?
In a vector database, the query is typically unstructured data such as text, image, audio, video, user profile, or molecule. This query is then vectorized and used to search for the nearest neighbors in the vector database. A vector database can contain billions of vectors, and the goal is to find the closest vectors to the query in the database. Once the closest vectors are identified, the original data is retrieved and used as the output.
Q4. What are the typical use cases for vector databases?
Retrieval augmented generation is one of them. Another very common example is recommender systems. Vector databases can also be used for text/semantic search, image/audio/video similarity search, question answering systems, molecular search, anomaly detetection and more. There are quite a few out there.
Q5. What are the main advantages of vector databases?
Vector databases offer semantic search capabilities, allowing users to search for information at a human level rather than just based on keywords. This enables users to understand information semantically, in addition to relying on labels. The result is often far more accurate and comprehensive search results.
Q6. What are the limitations?
Vector databases and vector search tools are not designed to handle relational data effectively. They are also very computationally intensive because of the large number of floating point numbers involved. These vectors require a lot of math operations, which can be a significant downside.
Q7. Who is already using vector databases in practice?
Many big and medium-sized tech companies rely on vector databases, which is why Zilliz is working on democratizing access to vector search technology. Milvus is an important part of that effort, as it aims to make vector search accessible at scales ranging from millions to billions of vectors.
Q8.Anything else you wish to add?
Nowadays, vector databases are mainly utilized for retrieval augmented generation. However, we will witness their widespread adoption in various applications in the next two to three years. The reason for this trend is quite simple – vectors are the future of search and represent unstructured data in a way that machines can understand.
To get the full lowdown on this topic, head over to https://zilliz.com/learn/what-is-vector-database
Frank Liu | Head of AI & ML
Frank Liu is the Head of AI & ML at Zilliz, with over eight years of industry experience in machine learning and hardware engineering. Before joining Zilliz, Frank co-founded an IoT startup based in Shanghai and worked as an ML Software Engineer at Yahoo in San Francisco. He presents at major industry events like the Open Source Summit and writes tech content for leading publications such as Towards Data Science and DZone. His passion for ML extends beyond the workplace; in his free time, he trains ML models and experiments with unique architectures. Frank holds MS and BS degrees in Electrical Engineering from Stanford University.