Visualising Vector Search: From Zero to Hero in Vector Search Algorithms | Milvus Community Meetup

Vector search can feel abstract - from high-dimensional embeddings and similarity metrics to ANN indexes, quantization, recall, filtering, and tuning. In this visual and practical session, Simon Hearne, Solutions Architect at Zilliz, takes the audience from zero to hero in vector search. Using intuitive visualizations and real-world benchmarks, Simon explains how vector search works, what changes at scale, and how to balance accuracy, latency, memory, and cost. This talk was recorded at the Milvus Meetup Paris, hosted by Criteo. The event brought together vector search practitioners to explore the algorithms, infrastructure, and production patterns behind modern AI-powered search. What you'll learn: - How keyword search and vector search complement each other - How embeddings turn text, images, and other data into vectors - How L2, cosine similarity, and dot product compare vectors - Why approximate nearest neighbor search is essential at scale - How HNSW, IVF, and DiskANN work - How quantization and dimensionality reduction reduce memory and cost - How Matryoshka Representation Learning and refinement help preserve recall - How filtering affects vector search performance - How to measure recall, precision, relevance, and business outcomes - What real-world benchmarks reveal about Q&A, code search, and agentic memory - How to choose between Milvus Lite, Milvus Standalone, distributed Milvus, and Zilliz Cloud Chapters: 0:00 Introduction 1:19 Why vector search exists 3:02 Models turn “stuff” into numbers 4:13 Picturing meaning: building an embedding model 5:52 What “similar” means: distance metrics 7:03 Exact search 7:52 Approximate nearest neighbour (ANN) 8:32 Recall and precision 9:40 HNSW: navigating a graph 11:14 IVF: partitioning the space 12:17 DiskANN: when RAM runs out 13:36 The size problem 14:36 Quantisation: SQ8, RaBitQ and PQ 17:49 Dimensionality reduction: PCA and MRL 20:33 Refinement: scan cheap, rescore precise 21:27 Filters are tricky 23:55 What Zilliz Cloud automates 25:13 When retrieval quietly fails 26:19 Measuring what you can’t see 28:30 Is the index worth it? Benchmarks 29:05 Q&A retrieval: the model sets the ceiling 30:16 Code search: grep vs index 32:51 Break-even on agentic memory 33:50 Where to run Milvus 34:38 Summary Find us on: Zilliz Website: https://zilliz.com Milvus Website: https://milvus.io Zilliz LinkedIn: https://www.linkedin.com/company/zilliz/ Milvus LinkedIn: https://linkedin.com/company/67143129 Milvus X: https://x.com/milvusio Milvus GitHub: https://github.com/milvus-io/milvus Try Zilliz Cloud: https://cloud.zilliz.com/signup