How Vector Databases Actually Work — It's Not Similarity Search

A vector database is not a similarity engine — it is an approximation engine. It is allowed to be wrong, on purpose, and every design choice inside (HNSW, quantization, DiskANN, IVF cells, filtering) falls out of that one idea. This is how they really search a hundred million vectors in milliseconds. Chapters: 0:00 The magic trick — what a vector DB claims to do 1:05 Why brute force dies (and the curse of dimensionality) 3:17 The ANN bargain — recall, latency, memory 5:15 HNSW part 1 — searching is walking, not scanning 6:33 HNSW part 2 — the express layers 8:39 Recall is a dial, not a property 9:41 The memory problem — quantization & disk 11:59 Filtering — the part nobody warns you about 14:00 What real systems actually use 15:29 Where it lives — RAG and retrieve-then-rerank 16:48 The one thing to remember What you'll learn: • Why exact nearest-neighbour search cannot work at scale — and what "approximate" really buys • How HNSW turns search into a greedy walk with express lanes (and why recall is YOUR dial) • Where you actually pay — recall vs latency vs memory — and how quantization/DiskANN move that cost #systemdesign #vectordatabase #softwareengineering The Hot Path — system design, explained in depth. New videos weekly. 🔔 Subscribe:    / @thehotpath   #systemdesign #softwareengineering #distributedsystems