The Minds Behind Vector Search and Similarity Indexing – 7 People Redefining Software

TL;DR

Vector search became practical by combining approximation, geometric hashing, graph navigation, compression, reusable libraries, and hardware acceleration. Indyk and Andoni advanced locality-sensitive hashing; Malkov co-created HNSW; Jégou and Douze developed product quantization and Faiss-era methods; Johnson optimized billion-scale GPU search; Lomet represents the classical database-indexing lineage that modern vector systems must integrate with.^[3][4][5]

Why you should read it anyway

Modern AI systems often turn text, images, audio, users, or products into vectors. Once information becomes a point in hundreds or thousands of dimensions, traditional exact-key indexes are no longer enough. The task changes from “find this key” to “find the points most similar to this point,” usually fast enough to search millions or billions of candidates interactively.

Imagine where Vector Search and Similarity Indexing would be without them

Without approximate similarity indexing, semantic search, large-scale recommendations, image matching, retrieval-augmented generation, and embedding-based discovery would be far more expensive. Systems could perform brute-force comparisons, but latency and infrastructure cost would rise dramatically as vector collections grew.

Time Estimate of how many years we would be hindered without them for human progress

Editorial counterfactual estimate: 4–8 years. Nearest-neighbor search is a fundamental computational problem, so multiple communities were pursuing solutions. The modern acceleration came from several ideas converging at the right time: approximation, HNSW graphs, vector compression, GPU kernels, and robust open libraries that made ANN search accessible to ordinary application developers.

The 7 people behind Vector Search and Similarity Indexing

1. Piotr Indyk

Why they matter: Indyk helped make approximate nearest-neighbor search in high-dimensional spaces practical through locality-sensitive hashing and related algorithms. His MIT profile highlights approximate nearest-neighbor search as a central research area and includes implementations based on LSH.^[1] The key shift was accepting controlled approximation in exchange for avoiding the prohibitive cost of exact search as dimensionality grows.

2. Alexandr Andoni

Why they matter: Andoni advanced locality-sensitive hashing with algorithms designed to improve the time-space tradeoffs of high-dimensional nearest-neighbor search.^[2] His work helped move LSH from a broad theoretical idea toward techniques with stronger performance guarantees. He represents the algorithmic branch of vector search in which randomized hashing tries to make similar vectors collide.

3. Yury Malkov

Why they matter: Malkov co-created Hierarchical Navigable Small World graphs, HNSW, one of the most influential modern approximate-nearest-neighbor indexes.^[3] HNSW organizes vectors into a multi-layer proximity graph and performs greedy navigation from coarse to fine layers. Its strong recall and practical speed helped make graph-based vector indexes a default choice in many modern systems.

4. Hervé Jégou

Why they matter: Jégou co-developed product quantization for compact nearest-neighbor search, splitting vectors into subspaces and representing them with short code combinations.^[4] He later co-authored the GPU similarity-search work behind Faiss.^[5] His contribution attacks both memory and computation: billions of high-dimensional vectors become more manageable when they can be compressed and compared efficiently.

5. Matthijs Douze

Why they matter: Douze co-authored product-quantization research with Jégou and later the Faiss GPU work with Jégou and Jeff Johnson.^[4][5] He represents the systems path from algorithm to reusable library. Vector search became broadly useful when developers could combine exact search, compressed indexes, inverted structures, graph methods, and GPU acceleration behind practical APIs.

6. Jeff Johnson

Why they matter: Johnson co-authored the billion-scale GPU similarity-search paper that demonstrated highly optimized k-selection and nearest-neighbor search on graphics processors.^[5] Faiss subsequently became a widely used library for vector indexing and similarity search.^[6] His role is hardware-aware systems optimization: exploit massive GPU parallelism without letting memory traffic and selection operations erase the benefit.

7. David Lomet

Why they matter: Lomet is deliberately the adjacent figure in this roster. His career centers on database indexing structures such as advanced B-tree variants rather than the invention of modern vector ANN algorithms.^[7] He matters because vector databases did not replace conventional indexing problems; they added a new retrieval geometry. Lomet represents the lineage of storage-engine indexing, page organization, and concurrency that vector systems must still coexist with.

How they each differ from one another

Indyk and Andoni represent locality-sensitive hashing and theoretical approximate search. Malkov represents navigable graph indexes. Jégou introduced major compression techniques and helped develop Faiss. Douze helped carry those algorithms into a reusable system. Johnson exploited GPUs for billion-scale similarity search. Lomet is not a vector-search inventor; he represents the classical database-indexing tradition that provides contrast and storage-engine context.

Final Take

Vector search is not the end of indexing history; it is a new branch of it. B-trees organize ordered keys, inverted indexes organize terms, and ANN structures organize neighborhoods in high-dimensional geometry. Modern data platforms increasingly need all three. The people behind vector search made semantic similarity cheap enough to become an everyday software primitive, which is why their work now sits directly beneath search, recommendation, and generative AI.

RESEARCH / PROVENANCE

Works Cited

7 SOURCES

1. 01
   Piotr Indyk — MIT CSAIL ↗people.csail.mit.edu
2. 02
   Alexandr Andoni — Locality-Sensitive Hashing ↗mit.edu
3. 03
   Malkov and Yashunin — Efficient and Robust Approximate Nearest Neighbor Search Using HNSW ↗arxiv.org
4. 04
   Jégou, Douze, and Schmid — Product Quantization for Nearest Neighbor Search ↗doi.org
5. 05
   Johnson, Douze, and Jégou — Billion-Scale Similarity Search with GPUs ↗arxiv.org
6. 06
   Faiss — The Faiss Library ↗arxiv.org
7. 07
   Microsoft Research — The Evolution of Effective B-Tree Page Organization and Techniques ↗microsoft.com

CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.