PART 5 · MODULE 4Storing AI knowledge
Vector Databases
Specialized databases for AI applications
Loading module...
Specialized databases for AI applications
Remember embeddings -- turning data into lists of numbers that capture meaning? Now imagine you have millions of those vectors. You need to search through them in milliseconds. That is what vector databases do.
Traditional databases are built for exact matches. "Find me the user with ID 12345" -- done. But AI needs similarity search: "Find me the documents most similar to this question." That requires a completely different approach.
Great at exact matches, filters, and joins. Struggles with "find me something similar to this."
WHERE name = 'John'Built for similarity. Finds the closest vectors to your query out of millions, in milliseconds.
NEAREST(query_vector, top_k=5)Vector space search — find the nearest neighbors
Checking every single vector against your query would be too slow with millions of entries. Vector databases use smart indexing to narrow down the search.
Imagine searching for a book about cooking. You would not check every single book in the library. Instead, you go to the "Cooking" section, then scan the shelves. Vector databases do the same thing -- they organize vectors into regions so they only need to check a small fraction to find the best matches.
Fully managed, easy to start
Open-source, feature-rich
Lightweight, developer-friendly
High-performance, scalable
Postgres extension, familiar
Search by meaning, not keywords. "How to request vacation" finds the PTO policy even if it never uses the word "vacation."
Find products, articles, or content similar to what a user already likes by comparing their embedding vectors.
The retrieval step in RAG uses vector search to find the most relevant documents for a user's question.
Vector databases are the memory layer of modern AI. They let models find relevant information instantly across millions of data points. Without them, every AI query would be like searching the entire internet from scratch -- possible in theory, impossibly slow in practice.