Vector Search & Semantic Retrieval

Vector search is what makes “find me things like this” work: semantic search, recommendations, and retrieval for LLMs. These posts go from the high-level idea down to the engineering: what vector databases actually do, how to implement a graph-based ANN index from a research paper, and how to make that index fast.

Read them in order if the topic is new to you, or jump to the one that matches where you’re stuck.

DateTitleRead
How Google Search Actually Works (It's Not What You Think) Learn what vector databases are, how semantic search works, and why ANN, DiskANN, and Vamana matter for modern retrieval systems.6 min
What I Got Wrong Implementing Graph-Based Vector Search A dev log about implementing a Vamana-style ANN index from the DiskANN paper, why my first version was slower than brute force, and what it taught me about reading algorithms with implementation in mind.9 min
How I Made My Vector Search Engine 16x Faster Without Changing the Algorithm A performance story from sembed-engine: the search visited the same number of nodes and returned the same answers, but became much faster by changing what the CPU had to do for every distance calculation.9 min
The Shape of Duplicate Detection Duplicate detection looks like a solved problem: use a hash set. A benchmark suite of 4050 measurements across finite batches and streaming workloads shows the fastest strategy can be 148x faster than a hash set, or 90,000x slower, depending on what you are deduplicating and what guarantees you need.16 min