Back to Videos
We Clustered 2 Million Product Embeddings. 40% of the Categories Were Wrong.
90
Retrieval & Search Research
Mixpeek Team
June 19, 2026
Summary
Embeddings are searchable but not readable: 768 or 1536 dimensions mean nothing to a human. This video clusters 2 million product embeddings, surfaces the structure hiding in vector space, and shows how clustering exposed that 40% of the human-assigned categories disagreed with what the products actually were. Explore: mixpeek.com/clusters
clusteringembeddingstaxonomyvector-searchdata-quality
About this video
Embeddings are searchable but not readable: 768 or 1536 dimensions mean nothing to a human. This video clusters 2 million product embeddings, surfaces the structure hiding in vector space, and shows how clustering exposed that 40% of the human-assigned categories disagreed with what the products actually were. Explore: mixpeek.com/clusters