Simon Hearne
LinkedIn
Search
VectorDB Zilliz Milvus talk

Vector Search, Visualised: Milvus Meetup Paris

A visual crash course in vector search, built around an ecommerce dress-finder: exact search, ANN indexes, quantisation, and the filters that quietly wreck your recall.
By
Published on

View Slides

About the talk

This is a refined version of my Vector Search, Visualised talk, rebuilt around a single ecommerce example: finding the right dress. A shopper searches for a Maison Lune dress for a summer wedding in Provence. Keyword search finds the brand but misses the occasion, vector search finds the occasion but blurs the brand, and hybrid search finds both.

From there, the talk follows that dress-finder from zero to production: how embeddings capture meaning, why exact search doesn't scale, what ANN indexes and quantisation trade away, and why a filter as ordinary as "size 38, under €150" can silently break your results. Vector search has no EXPLAIN, so the talk closes with how to measure what you can't see, from index recall all the way up to conversion and revenue per search.

What's covered

  • Why vector search exists. Keyword matches tokens, vectors match meaning. How models turn products into numbers, and what "similar" really means for a catalogue.
  • Exact vs. approximate search. Comparing a query to every product doesn't scale. ANN trades less than 10% recall for search that's over 100× faster and cheaper, framed as a triangle of speed, accuracy and cost.
  • ANN indexes, visualised. HNSW navigates a graph, IVF partitions the space, DiskANN reaches past RAM.
  • Making vectors smaller. Scalar, product and RaBitQ quantisation, PCA and Matryoshka embeddings, and refinement to buy precision back. Plus AUTOINDEX for when you'd rather not tune eight build parameters.
  • Filters and silent failure. Why harder filters destroy more of the graph, three ways out depending on selectivity, and how to catch the missing best match before your customers do.
  • Measure what you can't see. Golden queries scored against exact search on every deploy, and a layered metrics stack: index recall, relevance (NDCG, MRR), behaviour (CTR, zero-result searches) and business outcomes (conversion, return rate).
  • Is the index worth it? Break-even benchmarks for Q&A, code search and agent memory, and where to run it: Milvus Lite, Standalone or Distributed, or Zilliz Cloud.

The takeaway: spend recall on purpose. Check your model's ceiling, pick an index, shrink, buy back, filter wisely, and measure.

Presented at the Milvus Meetup Paris, hosted by Criteo, on 7 October 2026, alongside talks from Criteo on distributed vector search and Gorgias on RAG design patterns for product indexing.